Previous proposalML application lifecycle: recommendations for each stage
Next proposalWhat can software learn from robots and math
The shape of U
Submitted by Nishant Sinha (@ekshaks) via Zainab Bawa (@zainabbawa) on Sunday, 27 October 2019
Section: Crisp talk Technical level: Advanced Session type: Lecture Status: Confirmed & Scheduled
Tensors are the fundamental data structure for building modern machine learning programs and complex neural architectures. Unfortunately, the foundations of popular tensor libraries (numpy, tensorflow, pytorch) are hardly robust, e.g., tensor broadcasting rules are adhoc, and may cause surprising bugs. Further, the tensor library APIs expose low-level memory models to the developer, forcing them to continuously translate between their high-level mental models of data and low-level memory models. Moreover, absence of systematic ways to track shapes and perform ‘semantic’ transformations, forces them to guess the latent tensor shapes forever or add adhoc ‘shape comments’.
In short, we believe that developers trying to write deep learning programs/architectures from scratch or even trying to tweak existing model repositories and pre-trained models, are exposed to endless, unwanted ‘tensor’ misery.
In this talk, we will showcase our efforts at OffNote Labs to improve the developer experience when programming with tensors. In particular, we will discuss:
- The idea of naming dimensions of tensors and how named shapes can make tensor programming dramatically less painful.
- The tsalib library, which allows used named dimensions in Python 3.x programs with multiple backend libraries (numpy, tensorflow, pytorch, …).
- The tsanley library, which builds on tsalib, and helps catch tricky tensor shape errors at runtime and annotate existing programs with named shapes.
Nishant Sinha is an independent researcher and consultant at OffNote Labs, with broad experience in building deep learning systems (across text, vision and speech domains) and symbolic reasoning systems. Nishant helps companies understand and maneuver through the evolving deep learning/AI space and build IP, in-house teams and solutions that enable market leadership. He is also passionate about making cutting-edge research consumable and building tools that improve developer experience.
He received his Ph.D. from Carnegie Mellon University and B. Tech. in Computer Science from IIT Kharagpur.