Getting to know the elephant: Real-time debugging and visualization for deep learning

Shital Shah (Microsoft Research)

11:20am–12:00pm Thursday, September 26, 2019

Location: 1A 06/07

Data Science, Machine Learning, & AI

Average rating:

(4.00, 3 ratings)

Who is this presentation for?

Software engineers, machine learning engineers, and deep learning researchers

Level

Advanced

Description

While deep learning has been wildly successful at many tasks, our understanding of why these models work or don’t has been severely limited. Simultaneously, the area of model understanding and explaining its success of failures has been a fairly underrepresented area.

Shital Shah explores a new tool, TensorWatch, built at Microsoft Research to help and accelerate debugging and visualization for deep learning. This tool was designed from the ground up, rethinking fundamental primitives needed for such tasks and putting composability front and center. The framework is real time, interactive, and provides accelerated insights on how deep learning model works. Shital walks you through different aspects of workflows that machine learning practitioners frequently encounter. For example, quickly answering questions such as, what’s happening in the training, if the model architecture good enough, and why the model is failing for this given data. You’ll see a live interactive code running for demos.

Prerequisite knowledge

A basic understanding of deep learning, Python programming, Jupyter notebooks, statistics, and deep learning metrics

What you'll learn

Learn how to debug and visualize deep learning models, explore data, and investigate architecture issues with model

Shital Shah

Microsoft Research

Shital Shah is a principal research engineer at Microsoft Research AI. His interests include simulation, autonomous vehicles, robotics, deep learning, and reinforcement learning. At Microsoft, he’s architected, designed, and developed large-scale distributed machine learning systems. He conceived and lead the development of AirSim, a physically and visually realistic cross-platform simulator for AI research. Most recently, he developed TensorWatch, a new system for debugging and visualizing machine learning. He’s contributed in research and engineering in various roles at Microsoft, including technical lead, architect, engineering manager, and research engineer. Previously, he founded and led the team to design and develop distributed machine-learned clustering platform for web-scale data at Bing.