PipelineAI - Featured Startup SF '18

Chris Fregly: As a production-focused engineer, I'm surprised at how little attention is paid to production-izing ML and AI models.  This focus on high-performance model serving is a key difference between Scikit-Learn and TensorFlow, for example.  

Chris Fregly: The talk will highlight our recent R&D at PipelineAI.  We’re laser-focused on ML/AI model prediction performance including model transformations (ie. weight quantization) and runtime optimizations (ie. CPU vs. GPU).  We’ll show how decisions made during the model-training phase (ie. hyper-parameters like learning rate and decision tree depth) can affect performance and accuracy during the model-prediction phase in live production.

Q: Any additional background you would like our audience to know?

Chris Fregly: In essence, we’ve kept seeing the same problems over and over since 2010.  We realized that nobody is focused on “post-training” phases of the ML/AI pipeline.  By focusing on these phases, we’re surfacing a lot of optimizations that were previously left on the table.  Even small optimizations can lead to enormous performance and cost  benefits due to the scale of enterprise predictions (1,000,000’s of predictions per second) relative to training (100’s of training jobs per day).