Faculty Recruiting Support CICS

Engineering Approximate Computations

08 Nov
Friday, 11/08/2019 11:45am to 1:00pm
LGRC A311
Systems Lunch

There's a new ecosystem of applications that integrates machine learning into a variety of tasks. Typical domains have included image recognition and natural language processing. However, these techniques have also spread to computer systems domains, such as program compilation, resource scheduling, and database query optimization.

With the success of these systems, we must grapple with the reality that they model and compute with objects that are
inherently approximate -- real numbers (only computable up to a given precision), neural networks (only validated on a given dataset), and probabilistic computations (results only computable up to a given probability). This reality presents many questions about interpreting, debugging, validating, verifying, and optimizing these systems.

Guided by these questions, I'll present our work on systems for manipulating and reasoning about such approximate computations. I'll present our results on new programming systems for sound real-valued computation. I'll also present our work on the Lottery Ticket Hypothesis, a set of techniques for producing small trainable neural networks that are 10-20% of the size of standard architectures for MNIST, CIFAR10, and ImageNet. The promise of this latter work is not only faster inference and training, but also smaller neural networks that are more amenable to reasoning, such as verifying their robustness.

Faculty Host
: