Research

Computer Sciences and Information Technology

Title :

Model agnostic quantification of information transfer in deep neural networks

Area of research :

Computer Sciences and Information Technology

Principal Investigator :

Prof. Sumantra DuttaRoy, Indian Institute Of Technology (IIT) Delhi

Timeline Start Year :

2024

Timeline End Year :

2027

Contact info :

Equipments :

Details

Executive Summary :

The project aims to create mathematical tools to measure and evaluate the transferability of deep neural network (DNN)-based models for specific tasks. It will use time-frequency and topology-based methods to create an empirically easy-to-compute metric for assessing pre-trained acoustic models. The metric will analyze the trajectory growth of geometric objects passing through DNNs to link expressivity and generalization error in these models. The project also aims to study how these models can integrate spectral and/or temporal task-dependent information. The methods will be analyzed theoretically and empirically for various pre-trained models, downstream tasks, and modalities.

Co-PI:

Dr. Vinayak Abrol, Indraprastha Institute Of Information Technology, Delhi-110020

Total Budget (INR):

48,01,094

Organizations involved