Paper
Project Area
Model agnostic evaluation of transferability of pre-trained deep models

Funding Agency: IIITD-IITD

Details

Model agnostic evaluation of transferability of pre-trained deep models

Funding Agency: IIITD-IITD

Start Date: 2022-01-01    Duration: 2 years

PI: Vinayak Abrol

This project aims to develop mathematical tools to measure and evaluate the transferability of deep neural network (DNN) based models for a target task. To this aim various time-frequency (TF) and topology based methods will be explored to come up with an empirically easy to compute metric. Such a metric is aimed toward the practical assessment of a variety of different pre-trained acoustic models by running just a single pass through target data. We analyze the properties of our methods theoretically and also demonstrate their effectiveness empirically for diverse pre-trained models (supervised/unsupervised), downstream tasks (classification/regression), and modalities (vision/speech/audio).