log in  |  register  |  feedback?  |  help  |  web accessibility
PhD Proposal: Neural Networks as Databases - From data to model compression
Shishira R Maiya
Wednesday, August 14, 2024, 10:30 am-12:30 pm
  • You are subscribed to this talk through .
  • You are watching this talk through .
  • You are subscribed to this talk. (unsubscribe, watch)
  • You are watching this talk. (unwatch, subscribe)
  • You are not subscribed to this talk. (watch, subscribe)
Abstract

https://umd.zoom.us/my/shishira

 

The past decade has witnessed an exponential increase in data and the rise of deep learning systems to handle it. These systems primarily analyze, learn and interpret the data by excelling in exploiting patterns. This leads us to the question - can we use these patterns to efficiently store the data as well? Essentially this would mean moving from the paradigm of data compression towards model compression. This talk introduces and investigates a variety of methods in the realm of model compression, with the eventual goal of combining compression and understanding.
In the first part of this talk, I will introduce a framework for representing videos as compressed functions using Implicit Neural Representations (INRs). We have developed an algorithm that exploits both the spatial and temporal redundancies of a given video to achieve efficient compression and real-time decoding. Building on these compression results, I will discuss our work on transforming these INRs into useful representations. We introduce a hypernetwork-based INR system that enables us to use these compressed objects for downstream tasks such as video retrieval and understanding. Having established the paradigm of thinking about data compression as a model compression problem, I will then present a case study of model pruning and its effects on vision recognition models. Finally, I will delve into proposed future research directions. These include improving network design with random networks and achieving faster encoding with meta-learning for semantically accurate video representations.

Bio

Shishira R Maiya is a PhD student at the University of Maryland, College Park, advised by Prof. Abhinav Shrivastava. His research interests include network sparsity, video representation, compression and understanding.

This talk is organized by Migo Gui