We could not locate the page you were looking for.

Below we have generated a list of search results based on the page you were trying to reach.

404 Error
Peter Skomoroch is an entrepreneur, investor, and the former Head of Data Products at Workday and LinkedIn. He was Co-Founder and CEO of SkipFlag, a venture backed deep learning startup acquired by Workday in 2018. Peter is a senior executive with extensive experience building and running teams that develop products powered by data and machine learning. He was an early member of the data team at LinkedIn, the world's largest professional network with over 500 million members worldwide. As a Principal Data Scientist at LinkedIn, he led data science teams focused on reputation, search, inferred identity, and building data products. He was also the creator of LinkedIn Skills and Endorsements, one of the fastest growing new product features in LinkedIn's history. Before joining LinkedIn, Peter was Director of Analytics at Juice Analytics and a Senior Research Engineer at AOL Search. In a previous life, he developed price optimization models for Fortune 500 retailers, studied machine learning at MIT, and worked on Biodefense projects for DARPA and The Department of Defense. Peter has a B.S. in Mathematics and Physics from Brandeis University and research experience in Machine Learning and Neuroscience.
Dr. Sameer Singh is an Associate Professor of Computer Science at the University of California, Irvine (UCI). He is working primarily on robustness and interpretability of machine learning algorithms, along with models that reason with text and structure for natural language processing. Sameer was a postdoctoral researcher at the University of Washington and received his PhD from the University of Massachusetts, Amherst, during which he interned at Microsoft Research, Google Research, and Yahoo! Labs. He has received the NSF CAREER award, selected as a DARPA Riser, UCI ICS Mid-Career Excellence in research award, and the Hellman and the Noyce Faculty Fellowships. His group has received funding from Allen Institute for AI, Amazon, NSF, DARPA, Adobe Research, Hasso Plattner Institute, NEC, Base 11, and FICO. Sameer has published extensively at machine learning and natural language processing venues, including paper awards at KDD 2016, ACL 2018, EMNLP 2019, AKBC 2020, and ACL 2020.
There are few things I love more than cuddling up with an exciting new book. There are always more things I want to learn than time I have in the day, and I think books are such a fun, long-form way of engaging (one where I won’t be tempted to check Twitter partway through). This book roundup is a selection from the last few years of TWIML guests, counting only the ones related to ML/AI published in the past 10 years. We hope that some of their insights are useful to you! If you liked their book or want to hear more about them before taking the leap into longform writing, check out the accompanying podcast episode (linked on the guest’s name). (Note: These links are affiliate links, which means that ordering through them helps support our show!) Adversarial ML Generative Adversarial Learning: Architectures and Applications (2022), Jürgen Schmidhuber AI Ethics Sex, Race, and Robots: How to Be Human in the Age of AI (2019), Ayanna Howard Ethics and Data Science (2018), Hilary Mason AI Sci-Fi AI 2041: Ten Visions for Our Future (2021), Kai-Fu Lee AI Analysis AI Superpowers: China, Silicon Valley, And The New World Order (2018), Kai-Fu Lee Rebooting AI: Building Artificial Intelligence We Can Trust (2019), Gary Marcus Artificial Unintelligence: How Computers Misunderstand the World (The MIT Press) (2019), Meredith Broussard Complexity: A Guided Tour (2011), Melanie Mitchell Artificial Intelligence: A Guide for Thinking Humans (2019), Melanie Mitchell Career Insights My Journey into AI (2018), Kai-Fu Lee Build a Career in Data Science (2020), Jacqueline Nolis Computational Neuroscience The Computational Brain (2016), Terrence Sejnowski Computer Vision Large-Scale Visual Geo-Localization (Advances in Computer Vision and Pattern Recognition) (2016), Amir Zamir Image Understanding using Sparse Representations (2014), Pavan Turaga Visual Attributes (Advances in Computer Vision and Pattern Recognition) (2017), Devi Parikh Crowdsourcing in Computer Vision (Foundations and Trends(r) in Computer Graphics and Vision) (2016), Adriana Kovashka Riemannian Computing in Computer Vision (2015), Pavan Turaga Databases Machine Knowledge: Creation and Curation of Comprehensive Knowledge Bases (2021), Xin Luna Dong Big Data Integration (Synthesis Lectures on Data Management) (2015), Xin Luna Dong Deep Learning The Deep Learning Revolution (2016), Terrence Sejnowski Dive into Deep Learning (2021), Zachary Lipton Introduction to Machine Learning A Course in Machine Learning (2020), Hal Daume III Approaching (Almost) Any Machine Learning Problem (2020), Abhishek Thakur Building Machine Learning Powered Applications: Going from Idea to Product (2020), Emmanuel Ameisen ML Organization Data Driven (2015), Hilary Mason The AI Organization: Learn from Real Companies and Microsoft’s Journey How to Redefine Your Organization with AI (2019), David Carmona MLOps Effective Data Science Infrastructure: How to make data scientists productive (2022), Ville Tuulos Model Specifics An Introduction to Variational Autoencoders (Foundations and Trends(r) in Machine Learning) (2019), Max Welling NLP Linguistic Fundamentals for Natural Language Processing II: 100 Essentials from Semantics and Pragmatics (2013), Emily M. Bender Robotics What to Expect When You’re Expecting Robots (2021), Julie Shah The New Breed: What Our History with Animals Reveals about Our Future with Robots (2021), Kate Darling Software How To Kernel-based Approximation Methods Using Matlab (2015), Michael McCourt
This study group will follow the text Reinforcement Learning: An Introduction, Sutton and Barto, 2nd Edition, 2018 MIT Press. The group meets every Sunday at 10 am PT starting on April 11, 2021. The study group slack channel is #rl_2021 (you can join our slack community by clicking on "join us" at twimlai.com/community). Also, the study group leaders are looking for volunteers from the community to help lead the sessions so don’t hesitate to post a note in the channel if you’d like to volunteer.
Getting to know Andrea Banino Andrea’s background as a neuroscientist informed his work in deep learning. At DeepMind, Andrea’s research falls in the realm of Artificial General Intelligence, specifically memory, along with investigating ways to shape deep learning systems so they better mimic the human brain. “I think for us, we have a different sort of memory. We have very long-term memory, we have short-term memory. I argue that agents should be equipped with these sort of different timescale memory.” Introduction to Human and Machine Memory Human memory can be broadly categorized into two kinds: short-term, sometimes called “working” memory, and long-term memory. Working memory deals with immediate phenomena and manipulates it for other cognitive functions. Tasks like counting, drawing a still life, or putting together a puzzle, where you use recently encountered information to accomplish a goal involve working memory. Recurrent neural networks and LSTMs are working memory equivalent models which hold information “online” to solve a problem, and then usually let it go afterwards. Long-term memory can be further subdivided into episodic and semantic memory. Episodic memory, also called autobiographical memory, catalogues personal experiences and stores them in memories. This differs from semantic memory, which generally stores knowledge and concepts. For example, knowing what a bike looks like and what it does is semantic memory, while remembering a specific bike ride with a friend is stored in autobiographical memory. Andrea’s research background is in long-term episodic memory. There isn’t a really good long-term memory equivalent in ML models yet, but Andrea and his team have experimented with a few different arrangements. Long-Term Memory Models One interesting model Andrea explored is a memory-augmented neural network. This is a neural network connected to an external memory source, which allows it to write previous computations and reuse previous computation procedures when it encounters similar problems. Retrieval augmented models are another long-term memory equivalent that have the ability to look things up in their memory. However, unlike human minds, they don’t update or reconsolidate their memory based on new information; it’s just a constant cycle of check and replicate. Transformer models also seem promising as a substrate for long-term memory. However, Andrea notes that they have only been used to model language so far, so still limited data. One downside is that transformers are computation-heavy and difficult to scale, so it’s definitely an open area of research. Overfitting, in models and humans A common critique of deep learning models is that they have a tendency to overfit to their data set, and have difficulty generalizing as a result. While this is certainly an issue, Andrea brought up another really interesting point. Humans also memorialize, and there’s always the potential for overfitting as a person. One way evolution has helped prevent against that is by increasing the data set over time, as the set of human experiences our brains pull from increases as we age. Andrea mentioned that even humans are limited in our generalizability — limited by the data we take in. The link between memory and learning is that consistent experience enables generalization, so people take memories and use them to predict the future. In some ways, our brains aim to minimize uncertainty, and incorporating previously-known information about the environment helps us predict what’s going to happen in the future. Neural Network Navigation Task In 2018, Andrea and his colleagues published a paper that explored agent navigation via representation. The model they built was programmed to mimic the human hippocampus. To understand what this model looked like, Andrea explained the three types of cells in the hippocampus that work together for spatial analysis. Head direction cells fire when a person is facing a specific direction relative to their environment. Place cells on the other hand fire in a specific place, such as the town square or even one’s own bedroom. Grid cells fire in a hexagonal lattice format and are theorized to be the cells that allow us to calculate shortcuts. Andrea et al. trained a neural network with models that mimicked each of these three traits. Via experimentation, using methods like dropout and introducing noise, Andrea and his team were able to determine that all three artificial cell types were necessary for successful shortcut navigation. “We managed to make the representation emerge in our neural network, trained it to do path integration, a navigation task. And we proved that that was the only agent able to take a shortcut. So, it was an empirical paper to prove what the grid cells are for.” Ponder Net: an algorithm that prioritizes complexity Andrea’s most recent development is an algorithm called Ponder Net. As a general rule, the amount of computational power required for a neural network to make an inference increases as the size of a model’s input (like its feature dimensionality) increases, while the required computational power has no necessary relation to the complexity of a particular problem or prediction. By contrast, the amount of time it takes a human to solve a problem is directly related to the problem’s complexity. Ponder Net attempts to create neural networks that budget computational resources based on problem complexity. It does so with the introduction of a halting algorithm which helps to conserve inference time, so if the computer is confident about the solution, it can stop calculating early. How does it work? Pondering steps & the halting algorithm Ponder Net is based on previous work called adaptive computation time. Adaptive computation time (ACT) minimizes the number of pondering steps with a halting algorithm. In ACT, the algorithm finds a weighted average of the prediction, instead of a specific prediction. With Ponder Net, the probability of halting is found for each time step in the sequence. Andrea explained that the probability of halting is a Bernoulli random variable (think coin flip) which tells you the probability of halting at the current step, given that you have not halted at the previous step. From there, Ponder Net calculates a probability distribution by multiplying the probability at each time step in order to form a proper geometric distribution. Once we have that, the algorithm can then calculate the loss for each prediction in the sequence that we made. The loss can then be weighted by the probability where we altered that particular step. Andrea sees Ponder Net as a technique that can be applied in many different architectures, and he tested it on a number of different tasks. The team reported above state-of-the art performance, and that Ponder Net was able to succeed at extrapolation tests where traditional neural networks fail. Transformers & Reinforcement Learning Another project Andrea mentioned was a BERT-inspired combined transformer and LSTM algorithm he published in a recent paper. While LSTMs work great for reinforcement learning tasks, they do suffer from a recency bias which makes them less suited to long-term memory problems. Transformers perform better over a long string of information, however their reward system is more complicated and they have noisier gradients. Andrea’s algorithm applied a BERT masking training to features from a CNN which were then reconstructed. Figure 1 from CoBERL paper Combining the LSTM with a transformer reduced the size and increased the speed of the algorithm. Something clever Andrea did was letting the agent choose whether to use the LSTM alone or to combine with the transformer “I think there’s lots of stuff we can do to improve transformers and memory in general, in reinforcement learning, especially in relation to the length of the context that we can process.” Check out the podcast episode to learn more about Ponder Net, and reinforcement learning!
The issue of bias in AI was the subject of much discussion in the AI community last week. The publication of PULSE, a machine learning model by Duke University researchers, sparked a great deal of it. PULSE proposes a new approach to the image super-resolution problem, i.e. generating a faithful higher-resolution version of a low-resolution image. In short, PULSE works by using a novel technique to efficiently search space of high-resolution artificial images generated using a GAN and identify ones that are downscale to the low-resolution image. This is in contrast to previous approaches to solving this problem, which work by incrementally upscaling the low-resolution images and which are typically trained in a supervised manner with low- and high-resolution image pairs. The images identified by PULSE are higher resolution and more realistic than those produced by previous approaches, and without the latter’s characteristic blurring of detailed areas. However, what the community quickly identified was that the PULSE method didn’t work so well on non-white input images. An example using a low res image of President Obama was one of the first to make the rounds, and Robert Ness used a photo of me to create this example: I’m going to skip a recounting of the unfortunate Twitter firestorm that ensued following the model’s release. For that background, Khari Johnson provides a thoughtful recap over at VentureBeat, as does Andrey Kurenkov over at The Gradient. Rather, I’m going to riff a bit on the idea of where bias comes from in AI systems. Specifically, in today’s episode of the podcast featuring my discussion with AI Ethics researcher Deb Raji I note, “I don’t fully get why it’s so important to some people to distinguish between algorithms being biased and data sets being biased.” Bias in AI systems is a complex topic, and the idea that more diverse data sets are the only answer is an oversimplification. Even in the case of image super-resolution, one can imagine an approach based on the same underlying dataset that exhibits behavior that is less biased, such as by adding additional constraints to a loss or search function or otherwise weighing the types of errors we see here more heavily. See AI artist Mario Klingemann’s Twitter thread for his experiments in this direction. Not electing to consider robustness to dataset biases is a decision that the algorithm designer makes. All too often, the “decision” to trade accuracy with regards to a minority subgroup for better overall accuracy is an implicit one, made without sufficient consideration. But what if, as a community, our assessment of an AI system’s performance was expanded to consider notions of bias as a matter of course? Some in the research community choose to abdicate this responsibility, by taking the position that there is no inherent bias in AI algorithms and that it is the responsibility of the engineers who use these algorithms to collect better data. However, as a community, each of us, and especially those with influence, has a responsibility to ensure that technology is created mindfully, with an awareness of its impact. On this note, it’s important to ask the more fundamental question of whether a less biased version of a system like PULSE should even exist, and who might be harmed by its existence. See Meredith Whittaker’s tweet and my conversation with Abeba Birhane on Algorithmic Injustice and Relational Ethics for more on this. A full exploration of the many issues raised by the PULSE model is far beyond the scope of this article, but there are many great resources out there that might be helpful in better understanding these issues and confronting them in our work. First off there are the videos from the tutorial on Fairness Accountability Transparency and Ethics in Computer Vision presented by Timnit Gebru and Emily Denton. CVPR organizers regard this tutorial as “required viewing for us all.” Next, Rachel Thomas has composed a great list of AI ethics resources on the fast.ai blog. Check out her list and let us know what you find most helpful. Finally, there is our very own Ethics, Bias, and AI playlist of TWIML AI Podcast episodes. We’ll be adding my conversation with Deb to it, and it will continue to evolve as we explore these issues via the podcast. I'd love to hear your thoughts on this. (Thanks to Deb Raji for providing feedback and additional resources for this article!)
Stefan Lee is involved in a number of projects around emergent communication. To hear about more of them, in addition to the ViLBERT model, check out the full interview! TWiML Talk #358 with Stefan Lee. One of the major barriers keeping robots from full immersion in our daily lives, is the reality that meaningful human interaction is dependent on the interpretation of visual and linguistic communication. Humans (and most living creatures) rely on these signals to contextualize, operate and organize our actions in relation to the world around us. Robots are extremely limited in their capacity to translate visual and linguistic inputs, which is why it remains challenging to hold smooth conversation with a robot. Ensuring that robots and humans can understand each other sufficiently is at the center of Stefan Lee’s research. Stefan is an assistant professor at Oregon State University for the Electrical Engineering and Computer Science Department. Stefan, along with his research team, held a number of talks and presentations at NeurIPS 2019. One highlight of their recent work is the development of a model called ViLBERT (Vision and Language BERT), published in their paper, ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. BERT Models for Natural Language Processing BERT (Bidirectional Encoder Representations from Transformers) is Google’s popularized model that has revolutionized natural language processing (NLP). BERT is what’s called a language model, meaning it is trained to predict the future words in a sentence based on past words. Another example is GPT-2, which caused a controversy when released by OpenAI last year. Part of what makes them so special is that they are bidirectional, meaning they can contextualize language by reading data from both the left and right. This builds relationships between words and helps the model make more informed predictions about related words. They also pre-train models on large sets of unlabeled data using a transformer architecture, so data sequences don’t need to be processed in order, enabling parallel computations. Models like BERT work by taking “a large language corpus and they learn certain little things to build supervision from unlabeled data. They’ll mask out a few words and have them re-predict it based on the other linguistic context, or they’ll ask if a sentence follows another sentence in text.” Extending BERT to the Visual Domain ViLBERT is an extension of the BERT technique. To apply the BERT model to vision and language, the team worked with a data set called Conceptual Captions, composed of around 3 million images paired with alt-text. Their method is to mask out random parts of the image, and then ask the model to reconstruct the rest of the image given the associated alt-text. “Likewise, we’re asking, does this sentence match with this image or not? or masking out parts of the language and having it reconstruct from the image and the text. We’re designing this self-supervised multi-model task with this large weekly supervised data source.” Visual Grounding Stefan describes that “Most of natural language processing is just learning based on its association with other words. Likewise on the visual side, you’re learning to represent some sparse set of classes. Those classes often relate to specific nouns, but they don’t have a sense of closeness, so there’s no idea that the feature for a cat should be close to the feature for a tiger…The point of ViLBERT is to try to learn these associations between vision and language directly. This is something we usually call visual grounding of a word.” Humans do this naturally because we often have the inherent context to imagine a visual representation of something. For example, the words “wolf head” might bring up certain imagery of wolves, but machines lack the same visual associations. What Stephan is working towards with ViLBERT is to present the agent with something like, “red ball” and have it interpret that to reconstruct an image of a red ball. How Vilbert Works: Bounding Boxes and Co-Attentional Transformer In object detection, bounding boxes are used to describe the target area of the object being observed. For the purpose of ViLBERT, an image is dissected as a set of bounding boxes that are independent of each other. Each box is given a positional encoding that shows where the box was pulled from, but ultimately the order of the sequence is not important. BERT works in the same way where you have a sequence of words (a sentence) that are treated as independent inputs and given a positional embedding. “It’s just a set. It’s an unordered set. In fact, the actual input API for BERT looks the same in our model for the visual side and the linguistic side.” The bounding boxes are output by an R-CNN model trained on the Visual Genome. The R-CNN model “can produce quite a lot of bounding boxes and you can sample from it if you’d like to increase the randomness.” Something to note is that many of the bounding boxes are not well aligned and the data could come back as fairly noisy. “Sometimes you’ll have an object like a road…it doesn’t do a great job of honing in on specific objects.” While the model is not trained on the visuals from scratch, it still has to learn the association even when it might not be obvious. To train the model, certain parts of the image (bounding boxes) are removed, and the model is asked about alignment between the alt-text and the image. The distinction between BERT is that “In BERT, it’s a sentence and the next sentence, and you’re predicting whether one comes after the other, but in our case, it’s an image and a sentence, and we’re asking does this align or not? Is this actually a pair from conceptual captions?” At the end of the process, “what we get is a model that has built some representations that bridge between vision and language” which can then be fine-tuned to fit a variety of other tasks. Applications and Limitations Since the ViLBERT paper, the model has been applied to around a dozen different vision and language tasks. “You can pre-train this and then use it as a base to perform fairly well, fairly quickly, on a wide range of visual and language reasoning tasks.” In addition to fine-tuning for specific task types, you can adjust for specific types of data or data language relationships. One useful adaptation is for Visual Question and Answering (VQA) to help those who are visually impaired ask questions about the world around them, and receive descriptive answers in return. To modify the model for VQA, you could feed the questions as the text inputs and “train an output that predicts my subset of answers.” ViLBERT is pre-trained on a dataset of images and captions as the text input. For VQA, you would use the questions as the text input and have the model reconstruct answers as the output. While ViLBERT is a solid starting point for the field, Stefan notes that the grounding component of the research is still underdeveloped. For example, if the model is trained for VQA on a limited dataset like COCO images, there may be objects that are not accounted for because the machine never learned they existed. “One example that I like to show from a recent paper is that [COCO images] don’t have any guns. If we’ve fed this caption trained on COCO with an image of a man in a red hat with a red shirt holding a shotgun, and the caption is a man in a red hat and a red shirt holding a baseball bat, because he’s wearing what looks like a baseball uniform and he’s got something in his hands. It might as well be a baseball bat. If we talk back to these potential applications of helping people with visual impairment, that kind of mistake doesn’t seem justifiable.” Future Directions For Visual Grounding One related area of research that Stefan has started to branch into is the interpretation of motion. The problem with images is that it can often be difficult to distinguish between active behaviors and stationary behaviors. “For a long time in the community, grounding has been on static images, but there’s actually a lot of concepts that rely on motion, that rely on interaction to ground. I could give you a photo and you could tell me it looks like people are talking, but they could just be sitting there quietly as well.” There is less emphasis on the interaction, which is a key element not only to understanding communication, but for accuracy in reading a social situation. Machines are not yet able to catch on to these distinctions and it’s a growing area of interest for Stefan. For more on what Stefan is up to, be sure to check out the full interview with Stefan Lee for a deeper understanding of ViLBERT and the numerous projects the team is working on.
Sam Charrington: Hey, what's up everyone? This is Sam. A quick reminder that we've got a bunch of newly formed or forming study groups, including groups focused on Kaggle competitions and the fast.ai NLP and Deep Learning for Coders part one courses. It's not too late to join us, which you can do by visiting twimlai.com/community. Also, this week I'm at re:Invent and next week I'll be at NeurIPS. If you're at either event, please reach out. I'd love to connect. All right. This week on the podcast, I'm excited to share a series of shows recorded in Orlando during the Microsoft Ignite conference. Before we jump in, I'd like to thank Microsoft for their support of the show and their sponsorship of this series. Thanks to decades of breakthrough research and technology, Microsoft is making AI real for businesses with Azure AI, a set of services that span vision, speech, language processing, custom machine learning, and more. Millions of developers and data scientists around the world are using Azure AI to build innovative applications and machine learning models for their organizations, including 85% of the Fortune 100. Microsoft customers like Spotify, Lexmark, and Airbus, choose Azure AI because of its proven enterprise grade capabilities and innovations, wide range of developer tools and services and trusted approach. Stay tuned to learn how Microsoft is enabling developers, data scientists and MLOps and DevOps professionals across all skill levels to increase productivity, operationalize models at scale and innovate faster and more responsibly with Azure machine learning. Learn more at aka.ms/azureml. All right, onto the show! Erez Barak: [00:02:06] Thank you. Great to be here with you, Sam. Sam Charrington: [00:02:08] I'm super excited about this conversation. We will be diving into a topic that is generating a lot of excitement in the industry and that is Auto ML and the automation of the data science process. But before we dig into that, I'd love to hear how you got started working in ML and AI. Erez Barak: [00:02:30] It's a great question because I've been working with data for quite a while. And I think roughly about five to 10 years ago, it became apparent that the next chapter for anyone working with data has to weave itself through the AI world. The world of opportunity with AI is really only limited by the amount of data you have, the uniqueness of the data you have and the access you have to data. And once you're able to connect those two worlds, a lot of things like predictions, new insights, new directions, sort of come out of the woodwork. So seeing that opportunity, imagining that potential, has naturally led me to work with AI. I was lucky enough to join the Azure AI group, and there's really three focal areas within that group. One of them is machine learning. How do we enable data scientists of all skills to operate through the machine learning lifecycle, starting from the data to the training, to registering the models to putting them in productions and managing them, a process we call ML Ops. So just looking at that end to end and understanding how we enable others to really go through that process in a responsible trusted and known way has been a super exciting journey so far. Sam Charrington: [00:03:56] And so do you come at this primarily from a data science perspective, a research perspective, an engineering perspective? Or none of the above? Or all of the above? Erez Barak: [00:04:07] I'm actually going to go with all of the above. I think it'd be remiss to think that if you're  a data science perspective, and you're trying to build a product and really looking to build the right set of products for people to use as they go through their AI journey, you'd probably miss out on an aspect of it. If you just think about the engineering perspective, you'll probably end up with great info that doesn't align with any of the data science. So you really have to think between the two worlds and how one empowers the other. You really have to figure out where most data scientists of all skills need the help, want the help, are looking for tools and products and services on Azure to help them out, and I think that's the part I find most compelling. Sort of figuring that out and then really going deep where you landed, right? 'Cause if we end up building a new SDK, we're going to spend a whole lot of time with our data science customers, our data science internal teams and figure out, "Well, how should our SDK look like?" But if you're building something like Auto ML that's targeted not only at the deeper data scientist, but also the deeper rooted data professionals, you're going to spend some time with them and understand not only what they need, but also how that applies to the world of data science. Sam Charrington: [00:05:27] And what were you working on before Azure AI? Erez Barak: [00:05:31] So before Azure AI, in Microsoft, I worked for a team called Share Data, which really created a set of data platforms for our internal teams. And prior to joining Microsoft, I worked in the marketing automation space, at a company called Optify. and again the unique assets we were able to bring to the table as part of Optify in the world of marketing automations were always data based. We were always sort of looking at the data assets the marketers had and said, "what else can we get out of it?" Machine learning wasn't as prevalent at the time, but you could track back to a lot of what we did at that time and how machine learning would've helped if it was used on such a general basis. Sam Charrington: [00:06:12] Yeah, one of the first machine learning use cases that I worked with were with folks that were doing trying to do lead scoring and likelihood to buy, propensity to buy types of use cases. I mean that's been going on for a really long time. Erez Barak: [00:06:30] So we're on a podcast so you can't see me smiling, but we did a lot of work around building lead scoring...and heuristics and manual heuristics, and general heuristics, and heuristics that the customer could customize. And today, you've seen that to really evolve to a place where there's a lot of machine learning behind it. I mean, it's perfect for machine learning, right? You've got all this data. It's fresh. It's coming  in new. There's insights that are really hard to find out. Once you've start slicing and dicing it by regions or by size of customers, it gets even more interesting so all the makings for having machine learning really make it shine. Sam Charrington: [00:07:07] Yeah you are getting pretty excited I think. Erez Barak: [00:07:08] Oh, no, no, no. It's a sweet spot there. Yes. Sam Charrington: [00:07:12] Nice. You want to dive into talking about Auto ML? For the level of excitement and demand for Auto ML and enthusiasm that folks have for the topic, not to mention the amount of confusion that there is for the topic, I've probably not covered it nearly enough on the podcast. Certainly when I think of Auto ML, there's a long academic history behind the technical approaches that drive it. But it was really popularized for many with Google's Cloud Auto ML in 2018, and before that they had this New York Times PR win that was a New York Times article talking about how AI was going to create itself, and I think that contributed a lot to, 'for lack of a better term in this space', but then we see it all over the place. There are other approaches more focused on citizen data science. I'd love to just start with how you define Auto ML and what's your take on it as a space and its role and importance, that kind of thing. Erez Barak: [00:08:42] Yeah, I really relate to many of the things you touched on. So maybe I'll start - and this is true for many things we do in Azure AI but definitely for Auto ML - on your point around academic roots. Microsoft has this division called MSR, Microsoft Research, and it's really a set of researchers who look into bleeding edge topics and drive the world of research in different areas. And that is when we first got, in our team, introduced to Auto ML. So a subset of that team has been doing research around the Auto ML area for quite a few years. They've been looking at it, they've been thinking. It yes, I've heard the sentence, "AI making AI." That's definitely there. But when you start reading into it like what does it mean and to be honest, it means a lot of things to many people. It's quite overused. I'll be quite frank. There's no one industry standard definition that says, "Hmm, here's what Auto ML is." I can tell you what it is for us. I can tell you what it is for our customers. I can tell you where we're seeing it make a ton of impact. And it comes to using machine learning capabilities in order to help you, being the data scientist, create machine capabilities in a more efficient, in a more accurate, in a more structured fashion. Sam Charrington: [00:10:14] My reaction to that is that it's super high level. And it leaves the door open for all of this broad spectrum of definitions that you just talked about. For example, not to over index on what Google's been doing, but Cloud Auto ML Vision when it first came out was a way for folks to do vision cognitive services, but use some of their own data to tune it. Right? Which is a lot different. In fact, they caught a lot of flack from the academic Auto ML community because they totally redefined what that community had been working for for many years and started creating the confusion. Maybe a first question is, do you see it as being a broad spectrum of things or is it how do we even get to a definition that separates the personalized cognitive services trained with my own data versus this other set of things? Erez Barak: [00:11:30] I think you see it as more of that general sense, so I would say probably not. I see it as a much more concrete set of capabilities that adhere to a well known process. That actually is agreed upon across the industry. When you build a model, what do you do? You get data, you featurize that data. Once the features are in place, you choose a learner, you choose an algorithm. You train that algorithm with the data, creating a model. At that point, you want to evaluate the model, make sure it's accurate. You want to get some understanding of what are the underlining features that have most affected the model. And you want to make sure, in addition, that you can explain that model is not biased, you can explain that model is really fair towards all aspects of what it's looking at. That's a well-known process. I think there's no argument around that in the sort of the machine learning field that's sort of the end to end. Auto ML allows automating that process. So at its purest, you feed Auto ML the data and you get the rest for free if you may. Okay? that would be sort of where we're heading, where we want to be. And I think that's at the heart of Auto ML. So, where does the confusion start? I could claim that what we or others do for custom vision follows that path, and it does. I can also claim that some of what we do for custom vision is automated. And then there's  the short hop to say, "Well, therefore it is Auto ML." But I think that misses the general point of what we're trying to do with Auto ML. Custom vision is a great example where Auto ML can be leveraged. But Auto ML can be leveraged wherever that end to end process happens in machinery. Sam Charrington: [00:13:27] Nice. I like it. So maybe we can walk through that end to end process and talk about some of the key areas where automation is applied to contribute to Auto ML. Erez Barak: [00:13:44] So I'd like to start with featurization. And at the end of the day, we want an accurate model. A lot of that accuracy, a lot of the insights we can get, the predictions we can get, and the output we can get from any model is really hinged on how effective your featurization is. So many times you hear that, "Well, 80% of the time data scientists spend on data." Can I put a pin on, do you know where that number comes from? Oh of course. Everyone says that's the number, everyone repeats it. It's a self-fulfilling prophecy. I'm going to say 79% of it just to be sure. But I think it's more of an urban legend at that point. I am seeing customers who do spend that kind of percentages  I am seeing experiments rerun that take that amount of time. Generalizing that number is just too far now to do. Sam Charrington: [00:14:42] I was thinking about this recently, and wondering if there's some institute for data science that's been tracking this number over time. It would be interesting to see how it changes over time I think is the broader curiosity. Erez Barak: [00:14:55] It would. I should go figure that out. [laughs] So anyone who builds a model can quickly see the effect of featurization on the output. Now, a lot of what's done, when building features, can be automated. I would even venture to say that a part of it can be easily automated. Sam Charrington: [00:15:24] What are some examples? Erez Barak: [00:15:25] Some examples are like, "I want to take two columns and bring them together into one." "I want to change a date format to better align with the rest of my columns." And even a easy one, "I'd like to enhance my data with some public holiday data when I do my sales forecasting because that's really going to make it more accurate." So it's more data enhancement, but you definitely want to build features into your data to do that. So getting that right is key. Now start thinking of data sets that have many rows, but more importantly have many columns. Okay? And then the problem gets harder and harder. You want to try a lot more options. There's a lot more ways of featurizing the data. Some are more effective than others. Like we recently in Auto ML, have incorporated the BERT model into our auto featurization capability. Now that allows us to take text data we use for classification and quickly featurize it. It helps us featurize it in a way that requires less input data to come in for the model to be accurate. I think that's a great example of how deep and how far that can go. Sam Charrington: [00:16:40] You mentioned that getting that featurization right is key. To what extent is it an algorithmic methodological challenge versus computational challenge? If you can even separate these two. Meaning, there's this trade off between... Like we've got this catalog of recipes like combining columns and bending things and whatever that we can just throw at a data set that looks like it might fit. Versus more intelligent or selective application of techniques based on nuances whether pre-defined or learned about the data. Erez Barak: [00:17:28] So it extends on a few dimensions. I would say there are techniques. Some require more compute than others. Some are easier to get done. Some require a deeper integration with existing models like I mentioned BERT before, to be effective. But that's only one dimension. The other dimension is the fit of the data into a specific learner. So we don't call it experiments in machine learning for nothing. We experiment, we try. Okay? Nobody really knows exactly which features would affect the model in a proper way, would drive accuracy. So there's a lot of iteration and experimentation being done. Now think of this place where you have a lot of data, creating a lot of features and you want to try multiple learners, multiple algorithms if you may. And that becomes quickly quite a mundane process that automating can really, really help with. And then add on top of that, we're seeing more and more models created with just more and more features. The more features you have, the more nuanced you can get about describing your data. The more nuanced the model can get about predicting what's going to happen next, or we're now seeing models with millions and billions of features coming out. Now, Auto ML is not yet prepared to deal with the billion feature model, but we see that dimension extend. So extend compute, one, extend the number of iterations you would have, extend to the number of features you have. Now you got a problem that's quickly going to be referred to as mundane. Hard to do. Repetitive. Doesn't really require a lot of imagination. Automation just sounds perfect for that. So that's why one of the things we went after in the past, I'd say six to twelve months is how we get featurization to a place where you do a lot of auto featurization. Sam Charrington: [00:19:22] I'm trying to parse the extent to which, or whether, you agree with this dichotomy that I presented. You've got this mundane problem that if a human data scientist was doing would be just extremely iterative, and certainly one way of automating is to just do that iteration a lot quicker because the machine can do that. Another way of automating is... let's call it more intelligent approaches to navigating that feature space or that iteration space, and identifying through algorithmic techniques what are likely to be the right combinations of features as opposed to just throwing the kitchen sink at it and putting that in a bunch of loops. And certainly that's not a dichotomy, right? You do a bit of both. Can you elaborate on that trade off or the relationship between those two approaches? Is that even the right way to think about it or is that the wrong way to think about it? Erez Barak: [00:20:33] I think it's a definitely a way to think about it. I'm just thinking through that lens for a second. So I think you describe the brute force approach to it. On one side. The other side is how nuanced can you get about it? So what we know is you can get quite nuanced. There's things that are known to work, things that are not known to work. Things that work with a certain type of data set that don't work with another. Things that work with a certain type of data set combined with the learner that don't work with others. So as we build Auto ML, I talked about machine learning used to help with machine learning. We train a model to say, "Okay, in this kind of event, you might want to try this kind of combination first." Because if you're... I talked about the number of features, brute force is not an option. So we have have toto get a lot more nuanced about it, so what Auto ML does is given those conditions if you may, or those features for that model, it helps shape the right set of experiments before others. That's allowing you to get to a more accurate model faster. So I think that's one aspect of it. I think another aspect, which you may have touched on, and I think is really important throughout Auto ML, but definitely in featurization, is why people are excited about that. The next thing you are going to hear is that I want to see what you did. And you have to show what kind of features you used. And quickly follows is, "I want to change feature 950 out of the thousand features you gave me. And I want to add two more features at the end because I think they're important." That's where my innovation as a data scientist comes into play. So you've got to, and Auto ML allows you to do that, be able to open up that aspect and say, "Here's what I've come up with. Would you like to customize? Would you like to add? Would you like to remove?" Because that's where you as a data scientist shine and are able to innovate. Sam Charrington: [00:22:39] So we started with featurization. Next step is learner/model selection? Erez Barak: [00:22:45] I think it's probably the best next step to talk about. Yes. I think there's a lot of configuration that goes into this like how many iterations do I want to do?For instance. How accurate do I want to get? What defines accuracy? But those are more manual parameters we ask the user to add to it. But then automation again comes into play as learner selection. So putting Auto ML aside, what's going to happen? Build a set of features, choose a learner, one that I happen to know is really good for this kind of problem and try it out. See how accurate I get. If it doesn't work, but even if it works, you are going to try another. Try another few. Try a few options. Auto ML at the heart of it is what it does. Now, going to what we talked about in featurization, we don't take a brute force approach. We have a model that's been trained over millions of experiments, sort of knows what would be a good first choice given the data, given the type of features, given the type of outcome you want. What do we try first? Because people can't just run an endless number of iterations. It takes time, takes cost, and sort of takes the frankly it takes a lot of the ROI out of something you expect from Auto ML. So you want to get there as fast as possible based on learnings from the past. So what we've automated is that selection. Put in the data, set a number of iterations or not set them. We have a default number that goes in. And then start using the learners based on the environment we're seeing out there and choosing them out from that other model we've trained over time. By the way, that's a place where we really leaned on the outputs we got from MSR. That's a place where they, as they were defining Auto ML, as they were researching it, really went deep into, and really sort of created assets we were then able to leverage. A product sort of evolves over time and the technology evolves over time, but if I have to pick the most, or the deepest rooted area, we've looked at from MSR, it's definitely the ability to choose the right learner for the right job with a minimal amount of compute associated with it if you may. Sam Charrington: [00:24:59] And what are some of the core contributions of that research if you go to the layer deeper than that? Erez Barak: [00:25:10] Are you asking in context of choosing a model or in general? Sam Charrington: [00:25:13] Yeah, in the context of choosing a model. For example, as you described, what is essentially a learner, learning which model to use, that created a bunch of questions for me around like, "Okay how do you  represent this whole, what are the features of that model? And what is the structure of that model?" And I'm curious if that's something that came out of MSR or that was more from the productization and if there are specific things that came out of that MSR research that come to mind as being pivotal to the way you think about that process. Erez Barak: [00:25:57] So I recall the first version coming out of MSR wasn't really of the end to end product, but at the heart of it was this model that helps you pick learners as it relates to the type size of data you have and the type of target you have. This is where a lot of the research went into. This is where we publish papers around, "Well, which features matter when you choose that?" This is where MSR went and collected a lot of historical data around people running experiments and trained that model. So the basis at the heart of our earliest versions, we really leaned on MSR to get that model in place. We then added the workflow to it, the auto featurization I talked about, some other aspects we'll talk about in a minute, but at the heart of it, they did all that research to understand... Well, first train that model. Just grabbing the data. Sam Charrington: [00:26:54] And what does that model look like? Is it a single model? Is it relatively simple? Is it fairly complex? Is it some ensemble? Erez Barak: [00:27:06] I'll oversimplify a little bit, but it profiles your data. So it takes a profile of your data, it profiles your features, it takes a profile of your features. It looks at the kind of outcome you want to achieve. Am I doing time series forecasting here? I'm doing classification. I'm doing regression that really matters. And based on those features picks the first learner to go after. Then what it does is uses the result of that first iteration, which included all the features I'm talking about, but also now includes, "Hey, I also tried learner X and I got this result." And that helps it choose the next one. So what happens is you look at the base data you have, but you constantly have additional features that show you, "Well, what have I tried and what were the results?" And then the next learner gets picked based on that. And that gets you in a place where the more you iterate, the closer you get to that learner that gives you more accurate result. Sam Charrington: [00:28:14] So I'm hearing elements of both supervised learning. You have a bunch of experiments and the models that were chosen ultimately, but also elements of something more like simple reinforcement learning, contextual bandits, explore, exploit kind of things as well. Erez Barak: [00:28:37] It definitely does both. If I could just touch on one point, reinforcement learning, as it's defined, I wouldn't say we're doing reinforcement learning there. Saying that, we're definitely... every time we have an iteration going or every X times we have that, we do fine tune the training of the model to learn as it runs more and more. So I think reinforcement learning is a lot more reactive. But taking that as an analogy, we do sort of continuously collect more training data and then retrain the model that helps us choose better and better over time. Sam Charrington: [00:29:15] Interesting. So we've talked about a couple of these aspects of the process. Feature engineering, model selection, next is once you've identified the model, tuning hyper-parameters and optimization. Do you consider that its own step or is that a thing that you're doing all along? Or both? Erez Barak: [00:29:38] I consider it part of that uber process I talked about earlier. We're just delving into starting to use deep learning learner within Auto ML. So that's where we're also going to automate the parameter selection, hyper-parameter selection. A lot of the learners we have today are classic machine learning if you may, so that's where hyper-parameter tuning is not as applicable. But saying that, every time we see an opportunity like that, I think I mentioned earlier in our forecasting capabilities, we're now adding deep learning models. In order to make the forecasting more accurate, that's where that tuning will also be automated. Sam Charrington: [00:30:20] Okay, actually elaborate. I think we chatted about that pre-interview, but you mentioned that you're doing some stuff with TCN and Arema around times series forecasting. Can you elaborate on that? Erez Barak: [00:30:36] Yeah, so I talked about this process of choosing a learner. Now you also have to consider what is your possible set of learners you can choose from. And what we've added recently are sort of deep learning models or networks that actually are used within that process. So TCN and Arema are quite useful when doing times series forecasting. Really drive the accuracy based on the data you have. So we've now embedded those capabilities within our forecasting capability. Sam Charrington: [00:31:12] So when you say within forecasting, meaning a forecasting service that you're offering as opposed to within... Erez Barak: [00:31:21] No, let me clarify. There's three core use cases we support as part of Auto ML. One for classification, the other for regression, and the third for times series forecasting. So when I refer to that, I was referring more to that use case within Auto ML. Sam Charrington: [00:31:42] Got it. So in other words in the context of that forecasting use case, as opposed to building a system that is general and applying it to time series and using more generalized models, you're using now TCN and Arema as core to that, which are long-proven models for times series forecasting. Erez Barak: [00:32:07] Yeah, I would argue they're also a bit generalized, but in the context of forecasting. But let me tell you how we're thinking about it. There's generally applicable models. Now, we're seeing different use cases like in forecasting there are generally applicable models for that area, that are really useful in that area. That's sort of the current state we're in right now. And we want to add a lot more known generally applicable models to each area. In addition to that, sort of where we're heading and as I see this moving forward, more and more customers will want to add their own custom model. We've done forecasting for our manufacturing. We've tuned it to a place where it's just amazing for what we do because we know a lot more about our business than anyone else. We'd like to put that in the mix every time your Auto ML considers the best option. I think we're going to see- I'm already seeing a lot of that, sort of the 'bring your own model'. It makes sense. Sam Charrington: [00:33:06] That's an interesting extension to bring your own data, which was one of the first frontiers here. Erez Barak: [00:33:11] I mean you're coming in to a world now, it's not "Hey, there's no data science here. There's a lot of data science going on so I'm the data scientist. I've worked on this model for the past, you name it, weeks? Months? Years? And now this Auto ML is really going to help me be better? I don't think that's a claim we even want to make. I don't think that's a claim that's fair to make. The whole idea is find the user where they are. You have a custom model? Sure, let's plug that in. It's going to be considered with the rest in a fair and visible way, maybe with the auto featurization it even goes and becomes more accurate. Maybe you'll find out something else, you want to tune your model. Maybe you have five of those models, and you're not sure which one is best so you plug in all five. I think that's very much sort of where we're heading, plugging into an existing process that's already deep and rich wherever it lands. Sam Charrington: [00:34:07] The three areas that we've talked about, again featurization, model selection, and parameter tune or optimization are I think, what we tend to think of as the core of Auto ML. Do you also see it playing in the tail end of that process like the deployment after the model's deployed? There's certainly opportunities to automate there. A lot of that is very much related to dev ops and that kind of thing, but are there elements of that that are more like what we're talking about here? Erez Barak: [00:34:48] I think there's two steps, if you don't mind I'll talk about two steps before that. I think there's the evaluation of the model. Well, how accurate is it, right? But again you get into this world of iterations, right? So that's where automation is really helpful. That's one. The other is sort of the interpretation of the model. That's where automation really helps as well. So now, especially when I did a bunch of automation, I now want to make sure, "Well, which features really did affect this thing? Explain them to me. And work that into your automated processes. Did your process provide a fair set of data for my model to learn from? Does it represent all all genders properly? Does it represent all races properly? Does it represent all aspects of my problem, uses them in a fair way? Where do you see imbalance?" So I think automating those pieces are right before we jump into deployment, I think it's really mandatory when you do Auto ML to give that full picture. Otherwise, you're sort of creating the right set of tools, but I feel without doing that, you're sort of falling a bit short of providing everyone the right checks and balances to look at the work they're doing. So when I generalize the Auto ML process, I definitely include that. Back to your question on do I see deployment  playing there? To be honest, I'm not sure. I think definitely the way we evaluate success is we look at the models deployed with Auto ML or via Auto ML or that were created via Auto ML and are now deployed. We looked at their inferences. We look at their scoring, and we provide that view to the customer to assess the real value of their model. Automation there I think if I have to guess, yes. Automation will stretch there. Do I see it today? Can I call it that today? Not just yet. Sam Charrington: [00:36:54] Well, a lot of conversation  around this idea of deploying a model out into production, and thankfully I think we've convinced people that you can, it's not just deploy once and you're not thinking about it anymore. You have to monitor the performance of that model and there's a limited lifespan for most of the models that we're putting into production and then the next thing that folks get excited about is, "Well I can just see when my model falls out of tolerance and then auto-retrain..." It's one of these everyone's talking about it, few are actually doing it. it sounds like you're in agreement with that like we're not there yet at scale or no? Erez Barak: [00:37:42] So I think we often refer to that world as the world of ML ops. Machine learning operations in a more snappy way. I think there's a lot of automation there. If you look at automation, you do it dev ops for just code. I mean, forget machine learning code, but code, let alone models, is very much automation we need. I do think there're two separate loops that have clear interface points. Like deployed models, like maybe data about data drift. But they sort of move in different cycles at different speeds. So we're learning more about this but I suspect that iteration of training, improving accuracy, getting to a model where the data science team says, "Oh, this one's great. Let's use that." I suspect that's one cycle and frankly that's where we've been hyper-focused on automating with Auto ML. There's naturally another cycle of that, operations that we're sort of looking at automation opportunities with ML ops. Do they combine into one automation cycle? Hmm, I'm not sure. Sam Charrington: [00:38:58] But it does strike me that when for example, the decision "Do I retrain from scratch? Do I incrementally retrain? Do I start all the way over?" Maybe that decision could be driven by some patterns or characteristics in the nature of the drift in the performance shift that a model could be applied to. And then,  there're aspects of what we're thinking about and talking about as Auto ML that are applied to that dev ops-y part. Who knows? Erez Barak: [00:39:37] No, I'd say who knows. Then listening to you I'm thinking oh, to myself that while we sort of have a bit of a fixed mindset on the definition we'd definitely need to break through some of that and open up and see, "Well, what is it that we're hearing from the real world that should shape what we automate, how we automate and under which umbrella we put it?" I think, and you will notice, it's moving so fast, evolving so fast. I think we're just at the first step of it. Sam Charrington: [00:40:10] Yeah. A couple quick points that I wanted to ask about. Another couple areas that are generating excitement under this umbrella are neural architecture surge and neural evolution and techniques  like that. Are you doing anything in those domains? Erez Barak: [00:40:30] Again, we're incorporating some of those neural architectures into Auto ML today. I talked about our deeper roots with MSR and how they got us that first model. Our MSR team is very much looking deeper into those areas. They're not things that formulated just yet but the feeling is that the same concepts we put into Auto ML, or automated machine learning can be used there, can be automated there. I'm being a little vague because it is a little vague for us, but the feeling is that there is something there, and we're lucky enough to have the MSR arm that, when there's a feeling there's something there, some research starts to pan out, and they're thinking of different ideas there but to be frank, I don't have much to share at that point in terms of more specifics yet. Sam Charrington: [00:41:24] And my guess is we've been focused on this Auto ML as a set of platform capabilities that helps data scientists be more productive. There's a whole other aspect of Microsoft delivering cognitive services for vision, and other things where they're using Auto ML internally and where it's primarily deep learning based, and I can only imagine that they're throwing things like architecture surge and things like that at the problem. Erez Barak: [00:41:58] Yeah. So they do happen in many cases I think custom vision is a good example. We don't see the general patterns just yet and for the ones we do see, the means of automation haven't put out yet. So if I look at where we were with the Auto ML thinking probably a few years back is where that is right now. Meaning, "Oh, it's interesting. We know there's something there." The question is how we further evolve into something more specific. Sam Charrington: [00:42:30] Well, Erez, thanks so much for taking the time to chat with us about what you're up to. Great conversation and learned a ton. Thank you. Erez Barak: [00:42:38] Same here. Thanks for your time and the questions were great. Had a great time.
Bits & Bytes Microsoft open sources Bing vector search. The company published its vector search toolkit, Space Partition Tree and Graph (SPTAG) [Github], which provides tools for building, searching and serving large scale vector indexes. Intel makes progress toward optical neural networks. A new article on the Intel AI blog (which opens with a reference to TWIML Talk #267 guest Max Welling’s 2018 ICML keynote) describes research by Intel and UC Berkeley into new nanophotonic neural network architectures. A fault tolerant architecture is presented, which sacrifices accuracy to achieve greater robustness to manufacturing imprecision. Microsoft research demonstrates realistic speech with little labeled training data. Researchers have crafted an “almost unsupervised” text-to-speech model that can generate realistic speech using just 200 transcribed voice samples (about 20 minutes’ worth), together with additional unpaired speech and text data. Google deep learning model demonstrates promising results in detecting lung cancer. The system demonstrated the ability to detect lung cancer from low-dose chest computed tomography imagery, outperforming a panel of radiologists. Researchers trained the system on more than 42,000 CT scans. The resulting algorithms turned up 11% fewer false positives and 5% fewer false negatives than their human counterparts. Facebook open-sources Pythia for multimodal vision and language research. Pythia [Github] [arXiv] is a deep learning framework for vision and language multimodal research framework that helps researchers build, reproduce, and benchmark models. Pythia is built on PyTorch and designed for Visual Question Answering (VQA) research, and includes support for multitask learning and distributed training. Facebook unveils what secretive robotics division is working on. The company outlined some of the focus areas for its robotics research team, which include teaching robots to learn how to walk on their own, using curiosity to learn more effectively, and learning through tactile sensing. Dollars & Sense Algorithmia raises $25M Series B for its AI platform Icometrix, a provider of brain imaging AI solutions, has raised $18M Quadric, a startup developing a custom-designed chip and software suite for autonomous systems, has raised $15M in a funding Novi Labs, a developer of AI-driven unconventional well planning software, has raised $7M To receive the Bits & Bytes to your inbox, subscribe to our Newsletter.
Happy New Year! I've spent the week at CES in Las Vegas this week, checking out a bunch of exciting new technology. (And a bunch of not-so-exciting technology as well.) I'll be writing a bit more about my experiences at CES on the TWIML blog, but for now I'll simply state the obvious: AI was front and center at this year's show, with many interesting applications, spanning smart home and city to autonomous vehicles (using the term vehicle very broadly) to health tech and fitness tech. I focused on making videos this time around, and we'll be adding a bunch from the show to our CES 2019 playlist over on Youtube, so be sure to check that out and subscribe to our channel while you're there. In other news, we just wrapped up our AI Rewind 2018 series in which I discussed key trends from 2018 and predictions for 2019 with some of your favorite TWIML guests. This series was a bit of an experiment for us and we're excited to have received a lot of great feedback on it. If you've had a chance to check it out I'd love to hear your thoughts. Cheers, Sam P.S. We're always looking for sponsors to support our work with the podcast. If you think your company might benefit from TWIML sponsorship, I'd love your help getting connected to the right people. P.P.S. I'm planning a visit to the Bay Area for the week of January 21st. I've got a few open slots for briefings and meeting up with listeners. If you're interested in connecting give me a holler.   To receive this to your inbox, subscribe to our Newsletter.
Bits & Bytes IBM, Nvidia pair up on AI-optimized converged storage system.  IBM SpectrumAI with Nvidia DGX, is a converged system that combines a software-defined file system, all-flash storage, and Nvidia's DGX-1 GPU system. The storage system supports AI workloads and data tools such as TensorFlow, PyTorch, and Spark. Google announces Cloud TPU Pods, availability in alpha.  Google Cloud TPU Pods alpha are tightly-coupled supercomputers built with hundreds of Google’s custom Tensor Processing Unit (TPU) chips and dozens of host machines. Price/performance benchmarking shows a 27x speedup for nearly 40% lower cost in training a ResNet-50 network. MediaTek announces the Helio P90.  The Helio P90 system-on-chip (SoC) uses the company's APU 2.0 AI architecture. APU 2.0 is a leading fusion AI architecture designed by MediaTek can deliver a new level of AI experiences that are 4X more powerful than the Helio P70 and Helio P60 chipsets. Facebook open sources PyText for faster NLP development. Facebook has open sourced the PyText modeling framework for NLP experimentation and deployment. The library is built on PyTorch and supports use cases such as document classification, sequence tagging, semantic parsing, and multitask modeling. On scaling AI training. This interesting article from OpenAI proposes that the gradient noise scale metric can be used to predict parallelizability of training for a wide variety of tasks, and explores the relationship between gradient noise scale, batch size, and training speed. Dollars & Sense TechSee, Tel Aviv-based provider of AI-powered visual customer engagement solutions, has secured $16M in Series B funding Zesty.ai, an Oakland, CA-based AI startup, closed US$13M Series A financing Walmart Labs India, the product development division of the US retail giant, announced that it has acqui-hired AI and data analytics startup Int.ai Avnet, Inc announced that it will acquire Softweb Solutions, Inc., a software and AI company that provides software solutions for IoT Sign up for our Newsletter to receive this weekly to your inbox.
Bits & Bytes Google Cloud first to offer Tesla T4 GPUs. The NVIDIA Tesla T4 instances are currently available in alpha. They target ML inference, distributed training of models, and computer graphics workloads. Intel unveils Neural Compute Stick 2. Intel announced an update to the Movidius Neural Compute Stick focused on AI at the network edge. The Intel NCS 2 is based on the Intel Movidius Myriad X vision processing unit (VPU), offering more compute power and greater prototyping flexibility. Microsoft brings AI to Power BI. The new addition of AI features will bring image recognition, text analytics, and automated ML to Power BI. Integrated with Azure Machine Learning, Power BI will let users create ML models directly within the tool. Kasito launches second generation conversational AI platform. Kasisto’s updated conversational AI platform KAI will allow banks to launch new self-service features more easily. Linux Foundation launches Acumos platform for quick AI deployment. The Athena release of the Acumos AI Project aims to make building, sharing and deploying AI applications easy. In TWIML Talk #78, Mazin Gilbert talks about the goals and architecture of the platform. SnapLogic introduces self-service machine learning platform. SnapLogic announced the launch of SnapLogic Data Science, which combines data integration, data transformation, and machine learning training features in a single self-service platform. Elsevier creates AI-ready life sciences data platform. Entellect, a new cloud-based data platform aims to make life sciences R&D more efficient by enriching and harmonizing proprietary and external data and delivering it in an AI-ready environment. Dollars & Sense Standard Cognition, a provider of AI-powered autonomous checkout solutions, raised $40M in Series A funding Habana Labs, a startup in the AI processor space, announced it has raised $75M in Series B funding Apex.ai, which is building an operating system for self-driving vehicles, has raised $15.5M in Series A funding Microsoft announced its acquisition of a conversational AI and bot development startup XOXCO Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits & Bytes AWS introduces Amazon SageMaker Object2Vec. Entity embeddings are a powerful technique gaining traction among many ML users. With its new Object2Vec algorithm, AWS aims to make creating them more accessible for SageMaker users. Object2Vec is a highly customizable multi-purpose algorithm that can learn low-dimensional dense embeddings of high dimensional objects. Microsoft develops flexible AI system that can summarize the news. Summarization is a hard problem in NLP. Microsoft researchers have developed an AI framework that can reason relationships in “weakly structured” text, enabling it to outperform conventional models on a range of text summarization tasks. Google launches AI Hub and Kubeflow Pipelines. AI Hub provides a private, secure destination where enterprises can upload and share ML resources within their own organizations as well as access Google-produced content such as including pipelines, Jupyter notebooks, and TensorFlow modules. To support this, Kubeflow Pipelines will provide a way to compose, package, deploy, and manage ML workflows for reuse. The New York Times taps Google's AI to manage old photos. Google Cloud has teamed up with The New York Times to digitize and manage over 7 million archived photos and will use ML to find new insights from the old photos. OpenAI releases SpinningUp in Deep RL. For reinforcement learning fans and students in the community, this new educational resource from OpenAI provides a wealth of info and references to help users understand deep reinforcement learning. Would anyone be interested in a Deep RL course/study group based on this resource in early 2019? Dollars & Sense Engineer.ai announced that it has raised $29.5M Series A funding Aiden.ai, a London based AI analytics start-up, has raised a $1.6M Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits & Bytes ONNX Runtime for ML inference now in preview. Microsoft released a preview of the ONNX Runtime, a high-performance inference engine for Open Neural Network Exchange (ONNX) models. It is compatible with ONNX version 1.2 and comes in Python packages that support both CPU and GPU. Uber describes new platform for rapid Python ML development. Uber shared Michelangelo PyML, an extension to its Michelangelo platform providing for faster development and experimentation based on Docker containers. NYU and Facebook release cross-language NLU data set. As researchers look to increase the number of languages NLU systems can understand, gathering and annotating data in every language is a bottleneck. One alternative is to train a model on data in one language and then test that model in other languages. The Cross-Lingual Natural Language Inference (XNLI) data set advances this approach by providing that test data in languages. Malong researchers develop a technique to train deep neural networks. In this new paper, Malong introduces CurriculumNet, a training strategy leveraging curriculum learning to increase performance while decreasing noise when working on large sets of data. The code is now available on GitHub as well. Facebook launches Horizon reinforcement learning platform. Facebook has open-sourced Horizon, an end-to-end applied reinforcement learning platform. Unlike other open-source RL platforms focused on gameplay, Horizon targets real-world applications and is used at Facebook to optimize notifications, video streams, and chatbot suggestions. Google launches AdaNet for combining algorithms with AutoML. Google launched AdaNet, an open-source tool for automatically creating high-quality models based on neural architecture search and ensemble learning. Users can add their own model definitions to AdaNet using high-level TensorFlow APIs. Dollars & Sense People.ai announced that it has raised $30M in Series B funding led by Andreessen Horowitz DataRobot, a Boston-based automated ML company, raised $100M in Series D funding Syntiant Corp, an Irvine-based AI semiconductor company, raised $25M in Series B funding led by M12, Microsoft’s VC arm Oracle announced that it has acquired data management and AI solutions provider DataFox eSentire has acquired Seattle-based cybersecurity AI company Versive (formerly Context Relevant) AppZen, an AI auditing solutions provider, announced $35 million funding led by Lightspeed Venture Partners Validere, which provides an AI and IoT platform for oil and gas, raised $7m in seed funding Esperanto Technologies a hardware company focused on energy efficient systems for AI, ML, and DL, closed $58m Series B funding Conversica, offering conversational AI products for sales and marketing, announced it has secured a $31 million Series C funding Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits & Bytes Executive change at Google Cloud AI. Google Cloud AI head Dr. Fei-Fei Li has left the organization and will be returning to her professorship at Stanford. Dr. Andrew Moore, Dean of the School of Computer Science at CMU, will replace her by year end. Cisco unveils server for Artificial Intelligence and Machine Learning. Cisco has launched new servers aimed at speeding up deep learning workloads. Facebook's 'Rosetta' can extract text from a billion images daily. Facebook has developed a machine learning system called Rosetta for contextual extraction of text in images. The system supports image search and will also help Facebook identify inappropriate or harmful content. NVIDIA launches new data center inference. NVIDIA launched “TensorRT Hyperscale” offering inference acceleration for voice, video, image and recommendation services. The new platform features NVIDIA Tesla T4 GPUs based on the Turing architecture. Facebook developed an AI-based debugging tool. The tool, called “SapFix,” aims to help programmers by finding and fixing software bugs automatically. Google open sources ‘What-If Tool’ for code-free ML Experiments. Google’s AI research team has developed What-If Tool, a new TensorBoard feature allowing users to analyze an ML model without writing code. The tool offers a visual interface for exploring different model results. Dollars & Sense Syllable.ai, which offers a healthcare chat platform, raises $13.7M Integrate.ai, a Toronto-based AI software company, secures $30 million in Series A funding Microsoft announced the acquisition of San Francisco based Lobe, whose slick demo of code-free deep learning swept the Twitters a few months ago Deloitte announced its acquisition of Magnetic Media Online's artificial intelligence platform business. Magnetic is a marketing technology company headquartered in New York City Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
This video is a recap of our Fast.ai x TWIML Online Study Group. In this session, we review part one of lesson of seven, Resnets from Scratch. It’s not too late to join the study group. Just follow these simple steps: Sign up for the TWIML Online Meetup, noting fast.ai in the “What you hope to learn” box. Use the email invitation you’ll receive to join our Slack group. If you don’t receive it within a few minutes, check your spam folder. Once you’re in Slack, join the #fast_ai channel and hop over to #intros as well and introduce yourself. Use the link posted in the #meetup slack channel to add our events to your calendar.
Bits & Bytes Diffbot launches knowledge graph as-a-service. The startup, whose roots are in web scraping, applied machine learning, computer vision, and natural language processing to create a database of ‘all the knowledge of the Web,’ spanning over 10 billion entities and 1 trillion facts. Automatic transliteration helps Alexa find data across language barriers. Amazon researchers have developed a multilingual “named-entity transliteration system” to help Alexa overcome language barriers in multilingual environment. Oracle open sources GraphPipe for model deployment. Though Oracle has a strained relationship with open source, they recently released a new open source tool called GraphPipe, designed to simplify and standardize the deployment of machine learning models. Google turns datacenter cooling controls over to AI. Google was already using AI to optimize data center energy efficiency. Now they’ve handed over complete control of data center cooling to AI. Instead of humans implementing AI-generated recommendations, the system is now directly controlling data center cooling. IBM researchers propose ‘factsheets’ for AI transparency. Expanding on ideas like the Datasheets for Datasets paper I discussed previously, an IBM Research team has suggested a factsheet based approach for AI developers to ensure transparency. Facebook and NYU researchers speed up MRI scans with AI. Facebook announced the fastMRI project, in collaboration with NYU, which aims to apply AI to accelerate MRI scans by up to 10 times. Google releases Dopamine reinforcement learning research framework. Google announced the new TensorFlow-based framework, which aims to provide flexibility, stability, and reproducibility for new and experienced RL researchers. Baidu launches EZDL, a coding-free deep learning platform. Chinese firm Baidu EZDL, an online tool enabling anyone to build, design, and deploy models without writing code. Dollars & Sense Canvass Analytics, a Toronto-based provider of AI-enabled predictive analytics for IIOT, raised $5M in funding. Cloudalize, a cloud platform for running GPU-accelerated applications, has secured a €5 million funding round. Intel announced that it is buying Vertex.ai, a startup developing a platform-agnostic model suite, for an undisclosed amount. Zscaler, announced that it has acquired AI and ML technology and the development team of stealth security startup TrustPath. New Knowledge, an Austin-based cybersecurity company that protects corporations from covert, coordinated disinformation campaigns, raised $11M in Series A funding. Phrasee, a London based marketing technology company that uses AI to generate optimized marketing copy, closed a $4m Series A funding round. Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits & Bytes IBM Research presents 'DeepLocker,' AI-powered malware. IBM researchers have developed DeepLocker, a new breed of highly targeted and evasive attack tools powered by AI. The malware remains dormant until identifying its target through indicators like facial recognition, geolocation and voice recognition. The project is meant to raise awareness of the possibilities when AI is combined with current malware techniques. Microsoft Updates ML.NET Machine Learning Framework Microsoft has updated ML.NET, its cross-platform, open-source machine learning framework for .NET developers. This initial preview release of ML.NET enables simple tasks like classification and regression and brings the first draft of .NET APIs for training and inference. The framework can be extended to add support for popular ML libraries like TensorFlow, Accord.NET, and CNTK. Tesla dumps Nvidia, goes it alone on AI hardware. Tesla confirmed that it has developed its own processor for AutoPilot and other on-vehicle AI workloads. The company mentioned that the new processor is 10 times faster than the one it was using from Nvidia, and the linked article’s author estimates the company could save billions by developing its own. Doc.ai and Anthem to introduce health data trial powered by AI and blockchain. The companies have partnered to launch an AI data trial on the blockchain. doc.ai will attempt to identify models for predicting allergies based upon collected phenome (e.g., age, height, weight), exposome (e.g., exposure to weather/pollution based on location) and physiome (e.g., physical activity, daily steps) data. NetApp and NVIDIA launch new deep learning architecture. NetApp introduced the NetApp® ONTAP® AI architecture, powered by NVIDIA DGX™ supercomputers and NetApp AFF A800 cloud-connected all-flash storage to simplify, speed, and scale the deep learning data pipeline. Getty Images launches AI tool for media publishers. Getty Images launched Panels by Getty Images, a new artificial intelligence tool for media publishers that recommends the best visual content to accompany a news article. Dollars & Sense Malong Technologies, a Shenzhen, China-based computer vision startup, received a minority investment from Accenture Test.ai, which AI to the challenge of testing apps, has raised an $11 million Series A round led by Gradient Ventures Skyline AI, a New York based real estate investech firm using AI and data science, raised $18M in Series A funding DefinedCrowd, which provides crowd-sourced data for training AI, has raised a $11.8 million funding round led by Evolution Equity Partners Dbrain, whose platform links crowdworkers and data scientists via the blockchain, has raised a $3 million funding round Racetrack.ai, which applies AI to accelerating business sales, has raised $5 million in a pre-Series A round Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
This video is a recap of our July 2018 TWIML Online Meetup. In this month's community segment we look at the ongoing Fast.ai Study Group, the upcoming meetup presenter schedule, the recent Glow paper from the folks at OpenAI, and entity embeddings. In our presentation segment, Nicholas Teauge leads us in a discussion on the paper Quantum Machine Learning by Jacob Biamonte et al, which explores how to devise and implement concrete quantum software that outperforms classical computers on machine learning tasks. For links to the papers mentioned above and more information on this and previous meetups, or to get registered for upcoming meetups, visit twimlai.com/meetup! https://youtu.be/-ftniM7248I Paper: Quantum Machine Learning OpenAI Glow
Bits and Bytes This week news from Google I/O and Microsoft Build have dominated the news. Here are the highlights: Oracle rolling out AI applications for manufacturing. The applications leverage machine learning and AI to sift through large amounts of data from production environments to identify and trace issues from production through to customer delivery. IBM granted patent for AI-powered traffic management. The system would use computer vision powered cameras instead of timers to manage the flow of traffic. My friends over at SWIM are also doing interesting work in this area with one of their customers. Top Baidu AI executive stepping down. Top executive behind Baidu's artificial intelligence programs, Lu Qi, is stepping down. Lu is a former Microsoft executive and AI expert and has been responsible for day to day operations at Baidu's AI unit. Boston dynamics announces plans to sell SpotMini robot. The announcement came from Boston Dynamics founder Marc Raibert at the TC Sessions: Robotics conference at Berkeley. The robots are currently in pre-production but could available for sale come the middle of 2019. Researchers pair AI and drones to help manage agriculture. The University of South Australia system allows farmers to pinpoint areas that need more nutrients and water. This potentially improves crop outcomes and reduce resource mismanagement. Intel launches OpenVINO to accelerate computer vision development. The new toolkit, already in use at customers Agent Vi, Dahua, Dell, Current by GE, GE Healthcare, Hikvision, and Honeywell, includes three major APIs: The Deep Learning Deployment toolkit, a common deep learning inference toolkit, as well as optimized functions for OpenCV and OpenVX. Dollars & Sense Primal, an AI consumer and enterprise company, raises $2.3M BrainQ Technologies, a developer of AI to treat neuro-disorders, raises$8.8M in funding Motorleaf, a startup focused on data-driven insights for greenhouse and indoor operators, raises $2.85m Dialpad acquires TalkIQ to bring voice-driven AI to communications. Competitor 8x8 acquires MarianaIQ to strengthen AI capabilities as well. Oracle buys DataScience.com for data science platform Microsoft acquires Semantic Machines, to advance conversational AI Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits and Bytes Forgive the Facebook news bias here. There were a few interesting announcements from their F8 developer conference last week. Google’s Kubeflow brings machine learning support to Kubernetes. The open-source Kubeflow project for Kubernetes', Google's open-source container-orchestration system, has actually been around for a few months now and has seen strong interest in the open-source community. With the release of Kubeflow 0.1 it’s now a more robust option for building ML stacks on Kubernetes. Intel rolls out AI Builders Program for Enterprise AI partners. The program provides its members with resources to help bring their solutions to market on Intel AI technologies. These resources include technical enablement, marketing assistance, and in some cases investment. Facebook adds AI labs in Seattle and Pittsburgh. They’ve hired researchers from the University of Washington and Carnegie Mellon, causing some to worry that the AI talent shortage will ultimately be worsened by large companies poaching AI faculty. Facebook forms a special ethics team to prevent bias. The team spearheads a software system called Fairness Flow that is to help monitor implicit bias that might be unintentionally baked into production AI systems. Facebook announces PyTorch 1.0, a more unified AI framework. PyTorch 1.0 combines, Caffe2’s production-oriented capabilities and PyTorch's flexible research-focused design to provide an easier process for deploying of AI systems. Other companies have been quick to announce their compatibilities with the newly released software, like Microsoft and Google. Facebook Open Sources ELF OpenGo. Their AI bot, which is based on PyTorch and their ELF reinforcement learning platform, has successfully defeated 14 professional Go players as well as the strongest publicly available Go bot, LeelaZero. Facebook will be making both the trained model and the code used to create it open to the public. Dollars & Sense Suki, a startup creating a voice assistant for doctors, raises $20M Algolux Inc., a provider of machine-learning stacks for autonomous vision and imaging, raises $10M Passage AI, a provider of AI-powered conversational interfaces, raises$7.3 million MindBridge Analytics, a startup offering a FinTech autonomous auditing system, raises $8.4 Million BenchSci, a search engine to help researchers find antibody usage data, raises $10 million Synyi, a Chinese AI Medical Data startup, raises $15.7 million SoundHound, a competitor to Alexa and Google Assistant, raises$100M Humu, a behavioral-change software company, raises $30m Cisco to acquire Accompany, a company providing an AI-driven relationship intelligence platform for sales ServiceNow acquires Parlo, an artificial intelligence (AI) and natural language understanding (NLU) workforce solution Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits & Bytes Gartner pins “global AI business value” at $1.2 billion in 2018. Not sure what we’re supposed to do with that broadly-defined number beside putting it in our pitch decks, but I get it–AI just might be important ????. The firm describes three sources of AI value: customer experience enhancements, new revenues, cost reductions. Scientists in Europe call for new AI hub. Scientists in Europe pen an open letter outlining the motivations behind ELLIS–a new European AI hub. Nice analysis at The Guardian on this story and an EU commission call for €20 billion investment in the space to compete with the US and China. DeepCode launches to provide AI code copilot. Analogous to a grammar checker for code, the software connects to your GitHub repos and tells you how to fix problems in your Javascript, Java or Python code. AI helping discover new materials. Interesting article on recent Northwestern University research that applies AI to the task of creating new metal-glass hybrids. I did an interview on this topic recently so stay tuned for more. Google gives cloud credits to researchers. Awards are available to academic researchers in qualified countries. Awards are typically for $5k, but more can be requested via the application. Dollars & Sense Baidu has raised over $1.9 billion for a newly spun-off AI-powered financial services company Allegro.ai grabs $11 million to advance DL-as-a-Service platform Marble, a delivery robotics company raises $10 million iGenius, a Milan, Italy based startup, raises $7 million to make enterprise data easily queryable via natural language SalesHero raises $4.5 million seed to help reps keep Salesforce.com up to date Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits & Bytes Intel open sources nGraph neural network compiler. The newly open-sourced compiler, originally announced last summer and discussed on TWIML Talk #31, provides support for multiple deep learning frameworks while optimizing models for different hardware solutions. It supports six deep learning frameworks: TensorFlow, MXNet, neon, PyTorch, CNTK, and Caffe2. Google unveils augmented reality microscope. The prototype, which can detect cancer in real-time, was unveiled at an event organized by the American Association for Cancer Research. The new tool relays its predictions directly into the field of view of the user and has the ability to be retrofitted into existing microscopes. Google extends semantic language capabilities. Building on the hierarchical vector models at the heart of Gmails's Smart Reply feature, the new work extends these ideas by creating vectors for larger chunks of language such as full sentences and small paragraphs. The company published a paper on its Universal Sentence Encoder and launched the Semantic Experiences demonstration site. A pre-trained TensorFlow model was also released. IBM releases Adversarial Robustness Toolbox. The open-source software library aims to support researchers and developers in defending deep neural nets against adversarial attacks. The software, which currently works with TensorFlow and Keras, can assess a DNNs robustness, increase robustness as needed, and offer runtime detection of potential threats. MATLAB 2018a adds deep learning features. Many self-taught data scientists were initially exposed to MATLAB via Octave, the open source clone Andrew Ng used in his original Stanford machine learning online course. Well, the commercial software continues to evolve, with its latest version adding a host of new deep-learning related features including support for regression and bidirectional LSTMs, automatic validation of custom layers, and improved hardware support. Dollars & Sense Sword Health, a Portuguese medtech company, raises $4.6 million LawGeex, a contract review automation business, raises $12 million XpertSea, applying computer vision to aquaculture, raises $10 million Konux, a sensor and AI analytics startup, raises $20 million Citrine, materials data and AI platform, raises $8 million Eightfold.ai launches talent intelligence platform, closes $18 million round Voicera, the AI-powered productivity service, announces acquisition of Wrappup Adobe announces acquisition of voice technology business, Sayspring Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
In this month's community segment we chatted about explainability, Carlos Guestrin’s LIME paper, Europe’s attempt to ban “untrustworthy” AI systems and finally, Community member Nicolas Teague shares a blog post he wrote entitled "A Sight for Obscured Eye, Adversary, Optics, and Illusions,” which explores the parallels between computer vision adversarial examples & human vision optical illusions. In our presentation segment, Philosopie Group Inc. Director of AI, Chris Butler, joins us to discuss Trust in AI. Chris gives us an overview of a number of papers on the topic, including: Humans and Automation: Use, Misuse, Disuse, Abuse Trust in Automation: Integrating Empirical Evidence on Factors That Influence Trust Some Observations on Mental Models Overtrust of Robots in Emergency Evacuation Scenarios For links to the papers mentioned above and more information on this and previous meetups, or to get registered for upcoming meetups, visit [twimlai.com/meetup]!(/meetup) SUBSCRIBE AND TURN ON NOTIFICATIONS
Bits & Bytes Google develops AI that can pick out voices in a crowd. It is a deep learning audio-visual based model that uses both audio and video to isolate and enhance the targeted speaker while suppressing other sounds. The tech could be used in a wide range of applications from hearing aids to video conferencing. Microsoft halts sale of some enterprise AI tools over abuse fears. The tech giant is currently working with its internal Aether (AI and Ethics in Engineering and Research) Committee to review how AI tech could be used by its customers. There aren’t details on which applications have been ruled out but they've provided some insights into what issues they are prioritizing. Qualcomm’s launched two new chips to provide onboard AI processing to camera systems. Competing with AI inference silicon solutions like Intel Movidius and others, the AI edge system could be used in products like security cameras, drones, and robotics. Atos advances Quantum Learning Machine. The researchers have been able to successfully model quantum noise creating more realistic simulations. Not necessarily AI related but an interesting adjacent area. DimensionalMechanics updates NeoPulse framework. The new version includes updates to its NML modeling language for AI and new hyperparameter optimization features in its AI Studio. The company also raised an additional $1.25 million in an A-2 financing round. Dollars & Sense Juro, an AI startup for sales contracts, raises $2M Ocrolus, an AI company that analyzes financial documents, raises $4M Geoblink, a Spanish location intelligence startup, raises $6 million Mapillary, a startup developing a mapping system for autonomous vehicles, raises $15 million Xeeva, the procurement and sourcing software company, raises $40 million SleepScore Labs, a company providing sleep improvement systems, acquires Sleep.ai Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits and Bytes Apple hires Google’s AI head Google forms A.I. business unit. The latest in the AI talent wars, John Giannanderea, previously Google's chief of search and AI, was hired to run Apple’s “machine learning and A.I. strategy.” It’s an important victory for Apple who has lagged behind in AI. Google took the change as an opportunity to put AI into its own business unit under recent TWIML guest Jeff Dean. As the AI “arms race” intensifies, larger players are putting ever more resources into solidifying their positions. Last week we shared a similar story from Microsoft on its own reorg to better focus on AI. Researchers at MIT-IBM Watson AI Lab train models to recognize dynamic events. It’s easy for humans to recognize dynamic events, for example, opening a door, a book, or a bottle. MIT-IBM researchers hope to train models to recognize these types of dynamic events. They've released a Moments in Time dataset and are hosting a Moments in Time competition at CVPR. Note: I recently discussed similar work from the Univerisity of Montreal and startup Twenty Billion Neurons with its chief scientist Roland Memisevic. GridGain's newest release includes continuous learning framework. The company's in-memory computing framework based on Apache Ignite now includes machine learning and a multilayer perceptron (MLP) neural network that enables companies to run ML and deep learning algorithms against petabyte-scale operational datasets in real-time. Amazon SageMaker update. They’ve added support for more instance sizes and open sourced their MXNet and Tensorflow containers. The updated containers can be downloaded to support local development. Data scientist uses cloud ML to classify bowls of ramen. Nevermind hot dog/not hot dog... Data scientist Kenji Doi used Google Cloud AutoML Vision to successfully identify the exact shop each bowl is made at. A very impressive feat when you consider how similar the bowls of ramen actually look. Dollars and Sense Insider, an AI-enabled growth marketing platform, raises $11 million Comet.ml, a platform for managing AI projects, raises $2.3 million Audioburst, an AI-enabled audio search platform, raises $4.6 million from Samsung Conga to acquire, the contract discovery and analytics company Counselytics, to bolster AI strategy and document automation capabilities Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits and Bytes Last week I attended the GTC - GPU Technology Conference in San Jose. NVIDIA made quite a few announcements so you’ll see quite a few of those in this week’s news run down. Microsoft speeds neural net inference with Project Brainwave. Microsoft Research’s Project Brainwave uses Intel FPGAs to accelerate deep learning inference. The company reports that the system has been deployed for Bing, the search engine, resulting in 10x reductions in inference latency while accommodating a 10x increase in model size. Google launches text-to-speech service for developers. Cloud Text-to-Speech offers 32 different voices from 12 languages and variants. It includes a selection of voices built using WaveNet, a generative model for raw audio created by DeepMind. Microsoft reshuffles to bring more AI into products. With AI competition heating up industry-wide, Microsoft is looking to position itself as a leader in the AI solutions and developer markets. The company will now be split into “Experiences & Devices,” “Cloud + AI Platform,” and the existing branch of Microsoft Research. NVIDIA and Arm partner bring deep learning to IoT devices. NVIDIA and Arm will integrate the former’s open-source Deep Learning Accelerator architecture into the latter’s Project Trillium processors for machine learning inference. TensorFlow bumped to version 1.7 TensorFlow.js released.Version 1.7 of the framework moves Eager Mode, TF’s answer to PyTorch, into core. A GUI debugger is now offered in alpha as well. Support for NVIDIA’s TensorRT library for accelerated inference is included as well, among a bunch of other updates. Separately, the deeplearning.js project joins the TensorFlow family as TensorFlow.js. NVIDIA boosts deep learning platform. The NVIDIA Tesla V100 received at 2x memory boost to 32 GB plus the addition of GPU interconnect that enables up to 16 of the GPUs to communicate simultaneously. They also launched the DGX-2, an impressive machine targeting deep learning, capable of delivering two petaflops of computing power. YOLO v3 increases accuracy, humor. YOLO, short for You Only Look Once, is a popular image object detection system. The new version 3 offers minor improvements in accuracy explained in a very readable and quite funny research paper (PDF). Dollars and Sense Valohai, machine learning platform-as-a-service startup, raises $1.8 million Arraiy, a computer vision for the movie and TV industry, raises over $10M Scotty Labs, a startup working on remote-controlled driverless cars, raises $6 million with backing from Alphabet, Inc. Verbit, an AI transcription software startup, raises $11 million Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Healthcare applications of machine learning and AI have been in the news a bit more than usual recently, concurrent with the recent HiMSS conference in Las Vegas. HiMSS is a 45,000+ attendee conference dedicated to healthcare IT. Surprising no one, AI was a major factor at this year’s event. There was a whole subconference focused on ML & AI, plus a ton of AI-focused sessions in the regular conference and a good number of announcements by industry leaders and startups alike. I’ve only done a couple of healthcare-focused shows on the podcast so far, but I’m planning to dive into this area more deeply this year. Healthcare is arguably one the most promising–not to mention important–areas of AI application. Progress is being made across a good many areas, including: Radiology. Image-based diagnostics like radiology lend themselves to the application of deep learning. There are large amounts of labeled image data to work from and a degree of uniformity that's unmatched in many other vision applications. There’s been a raft of research papers on the application of deep learning to radiology and a lot of speculation about AI eventually replacing radiologists, but also strong arguments against this ever happening. Diagnostics. Radiology aside, machine learning and AI has the potential to help doctors make better diagnostic calls. One company that’s been active in this space is the startup Enlitic. The company–which at one time was lead by Fast.ai founder and former Kaggle president Jeremy Howard–wants to use deep learning to help make diagnostic calls, manage patient triage and screening programs, and identify high-level population health trends. Google Brain, Google DeepMind, and IBM Watson are all very active in this area as well, among others; the first of these recently published interesting research into the use of deep learning to predict cardiovascular diseases using retinal images, as opposed to more invasive blood tests. Health Monitoring. Machine learning is also driving health diagnostics and monitoring into the hands of consumers. Last year Apple unveiled a research study app that uses Apple Watch’s built-in heart rate monitor to collect data on irregular heartbeats and alert patients who may be experiencing atrial fibrillation. FirstBeat, a supplier to other fitness watch makers, uses machine learning to predict wearer’s stress levels and recovery times. I spoke with Ilkka Korhonen, the company’s vice president of technology about physiology-based models for fitness and training earlier this year. Personalized medicine. Personalized, or precision, medicine seeks to tailor medical interventions to the predicted response of the patient based on their genetics or other factors. Applications include selecting the best medicines for each patient and developing custom medications that target pathways based on an individual patient’s genetics. My interview with Brendan Frey of Toronto-based Deep Genomics explored a few of the opportunities in this space. Deep Genomics is working on “biologically accurate” artificial intelligence for developing new therapies. Electronic Health Records. The major EHR vendors–including Allscripts, Athenahealth, Cerner, eClinicalWorks, and Epic–all made announcements at HiMSS about ways that they would be incorporating AI into their products. Allscripts announced a partnership with Microsoft to develop an Azure-powered EHR, while Epic unveiled a partnership with Nuance to integrate their AI-powered virtual assistant into the Epic EHR workflow. Trump Administration advisor Jared Kushner even made an appearance advocating for greater EHR interoperability as a step towards applying AI, machine learning, and big data. Surgery. Researchers are beginning to incorporate AI into the planning and execution of a variety of surgical procedures. A variety of surgical scenarios have been explored, including burn care, limb transplants, craniofacial surgery, cancer treatment, and aesthetic (plastic) surgery. Of course, significant obstacles remain before we see AI fully integrated into healthcare delivery. Naturally, the barrier to releasing new products in healthcare is much higher than other industries since even small mistakes can have life-threatening consequences for patients. The techniques being applied now in research must be made more robust, a clear chain of accountability must be present, and justification for how why and how care decisions are made must be made clear. Improving robustness and performance will require time, a lot of data, and many rounds of testing. Increasing trust will further require new tools and techniques for explaining opaque algorithms like deep learning (the aforementioned Google research using retinal images provides a good example of this). We won’t see the autonomous robo-doctors of science fiction anytime soon, but machine learning and AI will undoubtedly play a significant role in the experience of healthcare consumers and providers in the years to come. Sign up for our Newsletter to receive this weekly to your inbox.
Bits and Bytes Google open sources exoplanet discovery AI. The project came out of a collaboration between Google Brain software engineer Chris Shallue and astrophysicist Andrew Vanderburg. The team was able to discover several new exoplanets and have now open-sourced their project to the public. I got a chance to talk with Chris Shallue about his work not too long ago, check out the show to learn more. Microsoft matches human performance translating news from Chinese to English. The research incorporated novel methods of training translation models including dual learning, deliberation, joint training and agreement regularization. Google's NSynth Super is an AI synth made of Raspberry Pis. The tool comes out of Magenta, Google’s creative AI applications project. The synthesizer uses open source AI software to generate new sounds. I talked with Doug Eck, the Magenta project lead, about his work on generative AI for music a little while back; give it a listen. Gluon models now deployable to AWS DeepLens. Gluon is an open source deep learning interface developed by AWS and Microsoft. It’s now deployable to AWS DeepLens instances for computer vision applications. Google open-sources the AI-powered tool for portrait mode on their Pixel devices. The tool uses semantic image segmentation to identify optimal focal areas or areas that need higher or lower exposure. Dollars & Sense SambaNova Systems, a start-up building computer processors and software for AI raises $56 million in funding led by Alphabet. Voci Technologies Incorporated, a provider of enterprise speech-to-text transcription and analytics, raises $8m in Series B funding. Percipient.ai, a provider of analytics for national security and now corporate security missions, raises $14.7M in Series A funding. Airspace Systems, Inc., a manufacturer of comprehensive drone defense systems, raised $20m in Series A funding. TaoData, a Chinese fintech startup, has raised $15.8 million in a series B round. Fractal Analytics, an AI solutions and analytics company, announced the acquisition of behavioral AI company Final Mile. L’Oréal announces the acquisition of ModiFace, an augmented reality and AI-powered beauty company. Avaya Holdings Corp. announced its acquisition of Spoken Communications, a Contact Center as a Service solutions application built on conversational artificial intelligence. Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits & Bytes Amazon to design its own AI chips for Alexa, Echo devices. This announcement follows similar moves made by rivals Apple and Google, both of which have developed custom AI silicon. Amazon, which reportedly has nearly 450 people on staff with chip expertise, sees custom AI chips as a way to make it's AI devices faster and more efficient. Google’s Cloud TPU AI accelerators now available to the public. Cloud TPUs are custom chips optimized for accelerating ML workloads in Tensorflow. Each boasts up to 180 teraflops of computing power and 64 gigabytes of high-bandwidth memory. Last week Google announced their beta availability via the Google Cloud. Cloud TPUs are available in limited quantities today and cost $6.50 / TPU-hour. At this cost, users can train a ResNet-50 neural network on ImageNet in less than a day for under $200. Finding pixie dust unavailable, Oracle sprinkles AI buzzword on cloud press release. The company applied "AI" to its Cloud Autonomous Services, including its Autonomous PaaS, and its Autonomous Database and Autonomous Data Warehouse products to make them "self-driving, self-securing and self-repairing" software. Oh boy! In other news, the company ran the same play for a suite of AI-powered finance applications. LG to introduce new AI tech for its smartphones. Following the launch of its ThinQ and DeepThinQ platforms earlier this year, as previously noted in this newsletter, LG will introduce new Voice AI and Vision AI features for its flagship V30 smartphone at the gigantic Mobile World Congress event next week. Applitools updates AI-powered visual software testing platform. I hadn't heard of this company before, but it's a pretty cool use case. The company released an update to its Applitools Eyes product, which is a tool for software development and test groups that allows them to ensure a visually consistent user experience as the application evolves. The company uses AI and computer vision techniques to detect changes to rendered web pages and applications, and report the ones that shouldn't be there. Dollars & Sense OWKIN, a company using transfer learning to accelerate drug discovery and development, closes $11m Series A financing. Ditto, a UK AI startup, raises £4 million to bring the expert system back via "software advisor" bots which aim to replicate human expertise and accountability. Palo Alto-based Uncomnon.co raises $18M in Series A funding for Uncommon IQ, its AI-powered talent marketplace. Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
Bits & Bytes Nest returns to the Google nest in AI push. Google is bringing device maker Nest back under its control as it fights Amazon and Apple for a foothold in the AI-enabled smart home market. Your smart watch could one day detect diabetes. A recent clinical study demonstrated that the Apple Watch, using its heart rate sensor paired with a machine learning algorithm, could detect diabetes with 85 percent accuracy. Photonic AI chip battle brewing. Last week I noted the $10 million fundraising by startup Lightelligence. Not to be outdone, Lightmatter, another startup with MIT roots going after the same light-powered AI chip market, announced its $11 million funding round. Facebook DensePose tools to re-skin videos in real time. The new Facebook AI Research system consists of DensePose-COCO, a human-annotated dataset which maps 2D images to 3D surfaces on the human body, and DensePose-RCNN, a new network architecture for classifying and localizing body surface parts from video frames. Version 0.2 of AWS Model Server for MXNet adds ONNX support. AWS updated its open source Model Server for Apache MXNet (MMS) library, which packages and serves deep learning models. New features include support for Open Neural Network Exchange (ONNX) models and the ability to publish operational metrics directly to Amazon CloudWatch. Dollars & Sense Paige.ai raises $25 million for cancer detection powered by computer vision Loris.ai, a Crisis Text Line spin-out, raises $2 million AdHive closes $5.5M ICO presale in 36 minutes for community-driven video advertising platform Boost AI raises $5M from Alliance Venture for a virtual AI-based custom agents
Bits and Bytes Interesting tidbits from recent news: Microsoft develops AI powered sketch artist. The new bot, based on recent GAN research, is capable of generating “drawings” from caption-like text descriptions. Applications for this technology include the arts, design, and perhaps at some point, police sketches. Overall very cool. IBM and Salesforce announce Watson + Einstein collaboration. The two tech giants are teaming up to integrate their two eponymously named, over-marketed, poorly understood machine learning products. Oh boy! Although it’s not immediately obvious in what ways Watson and Einstein are “combining”, Salesforce and IBM are making it clear that they are prioritizing AI and fleshing out their offerings. #SnarkLevelHigh Baidu grows AI research team. The new hires are Dr. Kenneth Church a pioneer in Natural Language Pioneering, Dr. Jun Huan a big data and data mining expert and Dr. Hui Xiong who specializes in data and knowledge engineering. Dating services firm Lunch Actually to launch ICO for Viola.AI. The dating service aims to not only match couples but also track their relationships, suggest date venues, remind them of new dates and advise them on relationship problems. Potentially a very interesting AI application, but one with tons of potential privacy implications. UC Berkeley & Facebook introduce House3D for reinforcement learning. The two teamed up to enable more robust intelligent agents by publishing a new dataset called “House3D”. House3D contains 45,622 3D scenes of houses, ranging from single-room studios to multi-storeyed houses equipped fully labeled 3D objects. In doing so, the groups aim to push RL research away towards focusing on tasks that more easily applicable to the real world. App claims to predict if an image will “go viral.” ParallelDots released the app with an open API that allows user to upload images then receive a “virality” score. It’s no secret that viral sharing is the dream of many marketers, so it’ll be interesting to see if this type of service could provide beneficial insights when planning ad campaigns. Amazon launched SageMaker BlazingText. BlazingText is an unsupervised learning algorithm for generating word2vec (see TT # 48) embeddings and is the latest addition to Amazon SageMaker’s suite of built-in algorithms. Deal Flow There seemed to be an abundance of deals last week: Smartphone-maker Coolpad has raised $300 million from Chinese property mogul Chen Hua-backed Power Sun Ventures to enhance its artificial intelligence capabilities. Understand.ai, a Karlsruhe, Germany-based machine learning startup for training and validation data in autonomous vehicles, raised $2.8 million in seed funding. C3 IoT, a provider whose software offerings include AI-for-IoT tools, announced a $100 million new round of financing. Data Nerds, a Canada-based developer of data products, raised $3m in Series A funding. Techcyte, Inc. closed a $4.3 million funding round to commercialize its digital pathology platform. Babblabs, a fresh start-up in advanced speech processing, announced today a Series Seed investment of $4 million. Owkin, a NYC-based predictive analytics company that utilizes transfer learning to accelerate drug discovery and development, raised $11m in Series A funding. Pony.ai, a year-old California-based self-driving car startup, announced it recently completed a $112 million Series A funding round. Smartsheet, that builds software for corporate process management, acquires business automation chatbot startup Converse.AI. Workday, the cloud HR and financials SaaS provider, buys SkipFlag to bolster machine learning capabilities. Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
This is a recap of the TWIML Online Meetup, held on Jan 16, 2018, we focus on the paper “Using deep learning and Google Street View to estimate the demographic makeup of neighborhoods across the United States” by Microsoft Research post doctoral researcher Timnit Gebru. We recap some of the major ML/AI Resolutions for 2018, community predictions for 2018, our favorite TWIML podcast episodes of 2017 and more. Thanks again to our presenter Timnit Gebru! Make sure you Like this video, and Subscribe to our channel below! https://youtu.be/pLdNx3SSxYo Full paper: "Using Deep Learning and Google Street View to Estimate the Demographic Makeup of Neighborhoods Across the United States" To register for the next meetup, visit twimlai.com/meetup Using Deep Learning to Estimate Demographic
Bits and Bytes Microsoft and Adaptive Biotechnologies want to decode the human immune system. The partners aim to create individual disease diagnostics, and ultimately a universal diagnostic, from a simple blood test using immunosequencing and machine learning. In other Microsoft news, the company is launching a $33 million AI hub in Taiwan. Microsoft will collaborate on AI research with a range of Taiwanese entities including government agencies, private sector, and academia. Intel brings AI tech to the Ferrari Challenge. Intel unveiled this deep computer vision application at CES, which uses fine-grained object detection to enable the personalization of race video streams. Stay tuned for my interview with the team’s lead data scientist in our upcoming CES coverage. Volkswagen and NVIDIA partner on autonomous vehicles. At the opposite end of the automotive and chip architecture spectra, the two companies announced plans to bring autonomous driving and AI-powered safety features to future cars, and unveiled the new I.D. Buzz concept which brings AI technology to the iconic VW MetroBus design. MediaTek launches cross-platform AI tech for consumer devices. Not to be left out, system-on-chip provider MediaTek is building out its NeuroPilot AI platform targeting consumer device manufacturers like Amazon, Belkin and Sony. The company also recently rolled out the Sensio 6-in-1 biosensor module that can track heart rate, blood pressure, peripheral oxygen saturation levels and more. DeepAR algorithm gives Amazon SageMaker new time-series capabilities. AWS added DeepAR support to its recently released SageMaker platform. The DeepAR algorithm is a supervised machine learning algorithm for forecasting using time-series data using recurrent neural networks (RNNs). Uber and Google explore “doubt” in deep AI systems. Interesting article the new crop of deep probabilistic programming tools including Uber’s Pyro and Columbia’s Edward. Unbabel nabs $23 million investment from Microsoft, Salesforce, Samsung for its translation software. Unbabel utilizes natural language processing, neural machine translation and quality estimation algorithms to bring greater accuracy to their translations. Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.
In this recap of the TWIML Online Meetup, held on Dec 13, 2017, we focus on the paper "Understanding Deep Learning Requires Rethinking Generalization" by Google Brain researchers Chiyuan Zhang, Samy Bengio and others. We also recap some of the major ML and AI advancements of the year, and take a look ahead to some of the key trends we’ll be following in 2018, such as deep reinforcement learning, capsule and schema networks and more. Thanks again to our presenter Bruno Gonçalves! Make sure you Like this video, and Subscribe to our channel above! https://youtu.be/mEYerIMYb5Q Full paper: Understanding Deep Learning Requires Rethinking Generalization To register for the next meetup, visit twimlai.com/meetup
A potentially interesting survey crossed the wires this week, and I while I’m bringing it up here, I do so with caveats, because the numbers seem a bit wonky. The survey, titled “Outlook on Artificial Intelligence in the Enterprise 2016” was published by Narrative Science, a “data storytelling” company that uses natural language generation to turn data into narratives. Narrative Science had help from the National Business Research Institute, a survey company that did the data collection for them. The headline of the survey announcement seems to be that 38% of those surveyed are already using AI technologies, while 56% of those that aren’t expect to do so by 2018. But, if that’s the case, then my math says that 73% of respondents’ organizations expect to have AI deployed by 2018, but the official report cites this number as 62%. Also, an infographic published by the same group says that only 24% of organizations surveyed are currently using AI, instead of the 38% quoted in their news release. This discrepancy could be due to the fact that a large percentage of organizations represented by the survey had more than one respondent, but it’s very confusing and I’d certainly expect more from a “data storytelling” company. Unless of course their press release and infographic where totally created by a generative AI, in which case I’m very impressed but also a bit horrified. Of course, the articles reporting on the survey don’t do anything to clear this up, with one of them reporting that 74% of organizations have already adopted AI. In any case, I feel we do need more data about enterprise adoption of AI, so some credible numbers here would be great but for now this ends up being just a cautionary tale about questioning your data. I have tweeted out to the company for clarification, and I’ll share whatever I find out. Subscribe: iTunes / Youtube / Spotify / RSS