Soumith Chintala is CTO of Thinking Machines Lab and a co-founder of PyTorch. His work spans machine-learning frameworks, image generation, open-source communities and robotics.[1][3][4]
Start with PyTorch: An Imperative Style, High-Performance Deep Learning Library →
Education
Links
Website
X 
GitHub LinkedIn Google Scholar Chief technology officer · Thinking Machines
Robotics research · New York University
Career
He spent eleven years at Meta before moving to Thinking Machines Lab, while continuing robotics work at NYU. Alongside PyTorch, he maintained the earlier Torch-7 framework and created convnet-benchmarks, a suite used to compare deep-learning implementations.[1]
Contributions
PyTorch makes a model’s computation a regular Python program under the developer’s control. That design lets researchers inspect intermediate values and debug with familiar tools while the runtime handles efficient execution on processors such as GPUs. The 2019 paper, which Chintala coauthored with a large team, explains how usability and runtime performance fit together rather than treating them as mutually exclusive goals.[2]
Learning representations through image generation
DCGAN, coauthored with Alec Radford and Luke Metz, investigates learning image representations without category labels. A convolutional generator and discriminator are trained together, and the researchers test whether their learned features transfer to other tasks. The paper’s contribution is a constrained architecture and empirical study of those representations.[3]
Ideas
Chintala emphasizes fast iteration and simpler systems. He treats open-source work as community design, with useful tools and feedback from the people applying them.[1]
Tools, writing & talks
Writing, research & talks
Start here
The design paper explains how ordinary Python control and an efficient runtime support deep-learning experimentation.
Start with the design principles and trace one small training example alongside the paper. Basic Python and neural-network vocabulary help; the paper explains the system rather than providing an installation tutorial.
A study of convolutional generator and discriminator networks as learned image representations.
Co-author[3]