Bits & Bytes – 11.14.17

1024 1024 The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)

Bits & bytes

  • AT&T launches open-source AI platform. Acumos—an extensible, framework-independent platform for delivering enterprise ML & AI solutions—was recently open-sourced under the auspices of the Linux Foundation. Speaking to AT&T exec Mazin Gilbert about the project, Acumos reminds me a bit of Intel’s mothballed Trusted Analytics Platform (TAP) project in complexity, scope and ambition. I wish them luck, but call me a skeptic here.
  • AI pioneer Geoffrey Hinton outlines new advance that requires less data. Hinton has been hinting (SWIDT?) at the challenges of deep learning and the need for something better for some time now. Turns out he had something up his sleeve: capsule networks, a new type of neural network that’s shown promising results on simple datasets.
  • Generating photorealistic images of fake celebrities. NVIDIA researchers use a technique called “progressive growing” with GANs to produce strikingly real-looking images based on a training dataset of celebrity photographs.
  • Uber AI Labs open sources Pyro. Uber AI has published Pyro, a tool for deep probabilistic modeling that seeks to unify the best of modern deep learning and Bayesian modeling. Would anyone be interested in demoing this, or seeing a demo of this, at an upcoming meetup?
  • Salesforce Introduces myEinstein. Salesforce has with Einstein a problem like IBM’s does with Watson. They’ve made it a broad brand that they apply to everything AI and which, consequently, has no specific meaning. That said, the feature they’ve announced, the ability for non-data-scientist Salesforce admins to add predictive fields to Salesforce layouts, actually sounds pretty useful if it works as advertised.

Sign up for our Newsletter to receive the Bits & Bytes weekly to your inbox.

Leave a Reply

Your email address will not be published.