Deep Signal Processing for Machine Learning Models
Abstract
We propose work on strengthening the connection between classical signal processing and modern deep learning to improve the representation quality, efficiency, and scalability of the next generation of machine learning models for handling long sequences of data. Sequential data, which is common to domains such as language, genomics, healthcare, finance, and audio has complex structure and long-range dependencies. Current deep learning architectures struggle to capture this structure. We propose a technical approach to leverage ideas from signal processing for modeling time-varying signals that have been relatively under-explored for machine learning. Our group#s past theoretical and applied work adapting state space models to deep learning has enabled improved models for a varietyof downstream tasks, but much work remains to adapt other signal processing primitives to deep learning. We focus on three aspects:developing deep signal processing architectures for modeling long sequences, developing efficient algorithms for these architectures, and understanding the data scaling laws for these architectures. If successful, the proposed research will allow users to better model signals over sequential data in critical application domains.** Approved for Public Release **
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 24, 2023
- Source ID
- N000142312633
Entities
People
- Christopher Ré
Organizations
- Office of Naval Research
- Stanford University
- United States Navy