Model Development and Training

Data science is a scientific endeavor and requires experimentation in order to identify and optimize a predictive model. Managing these experiments via a software tool or system allows data scientists to get experiment parameters and results out of spreadsheets (or worse, file names, post-it notes, or their heads) and into a tracking system aiding repeatability, collaboration, optimization, and automation. The organizations we’ve profiled have all built robust experiment management and model development features into their platforms to allow data scientists to delegate repetitive aspects of model training and optimization to machines, freeing their time to focus on high-value creative activities. Here are some of the most common capabilities of these systems:

• AutoML (Automated Machine Learning);
• Support for a variety of development approaches and tools;
• Model marketplace or store;
• Model repository for versioning models;
• Model training (regular and distributed);
• Model debugging
• Experiment management and visualization;
• Deep learning and reinforcement learning support;
• Hyperparameter tuning and optimization.