| PhD Seminar

Name of the Speaker: Mr. Dara Nagaraju (EE15D007)
Guide: Prof. Nitin Chandrachoodan
Venue: ESB-234 (Malaviya Hall)
Date/Time: 30th January 2024 (Tuesday), 11:00 AM
Title: Optimizing convolutional neural networks for resource-limited embedded devices.

Abstract

Convolutional neural networks (CNN) are used in several computer vision tasks such as image classification, localization, segmentation, and super-resolution. These networks contain a large number of multiply and accumulate operations (MAC) and parameters making them difficult to implement on resource-limited devices.

In this work, certain mathematical properties of the trained weight models/matrices are studied to identify redundancies and similarities of features which is used to decrease the computations and the parameters for the inference phase.

In the permuted features (PF) approach, we showed that the weights in FC layers can be modeled as permutations of a common sequence with minimal impact on accuracy. The PF approach reduces parameters from 1.3X to 36X (median: 4.45X). We also discuss how to use the PF approach on resource-limited devices with limited on-chip memory. In the QRCNN framework, graceful degradation of performance is shown while dropping features in the convolutional layers without retraining. The QRCNN method achieves a 1.22 -- 2.28X reduction in MAC computations (median: 1.575X). We also present the speedup achieved using this approach on Raspberry Pi 4B (embedded platform).