Computer Sciences and Information Technology
Title : | Hardware Aware Frame-Theoretic Sparse Neural Network for Resource-constrained Platform |
Area of research : | Computer Sciences and Information Technology |
Principal Investigator : | Mr. Pradip Sasmal, Indian Institute Of Technology (IIT) Jodhpur, Rajasthan |
Timeline Start Year : | 2024 |
Timeline End Year : | 2026 |
Contact info : | psasmal@iitj.ac.in |
Equipments : | GPU
Workstation |
Details
Executive Summary : | Deep Learning (DL) has shown immense potential in various applications, including self-driving cars, healthcare decision-making, and signal processing. However, it is crucial to understand the limitations of DL methods and establish a strong mathematical foundation. DL algorithms are power and memory-hungry, making them difficult to deploy on resource-constrained platforms. Training and running inference on these models are computationally expensive, and most run on large data centers with clusters of CPUs and GPUs, requiring a large power supply. Researchers are seeking low computational methods to reduce model size, such as pruning, which reduces the size of a sub-network that can perform as well as the original neural network (NN). However, most existing pruning methods lack mathematical explanations and depend heavily on simulations. The lottery ticket hypothesis and Caratheodory lemma suggest that there exists a sub-neural network that matches the test accuracy of the original neural network. Most existing pruning methods are heuristic and only a few meet available resources on the platform.
The objective of this project is to build hardware-aware low computational robust mathematical pruning methods for selecting suitable sparse sub-neural networks that can perform as good as the initial full network and are deployable on resource-constrained edge devices. In this context, the authors propose results from frame theory to develop a novel pruning method that can be deployed in mobile devices. In order to make India self-reliant on AI, it is essential to develop novel pruning methods for sustainable AI development in India. |
Total Budget (INR): | 18,29,322 |
Organizations involved