Under way

Evaluation of Stochastic Computing for Efficient Deep Neural Network Inference

Type of work:

Master Thesis

Assignment:
  • Literature review: Basic research about stochastic computing and hardware accelerators for machine learning.
  • Development: Developing a software framework for basic computation of neural networks using stochastic computing.
  • Integration: integrating the developed framework with Pytorch.
  • Evaluation:  Investigating the impact of using stochastic computing on the accuracy of neural network, throughput, and hardware implementation.
  • Documentation: Writing thesis report consisting of framework implementation, setup, results and conclusion
 
Skills:
  • Basic understanding of deep neural networks and stochastic computing
  • C++/CUDA
  • Pytorch
 
Background:

Deep neural networks (DNNs) require substantial memory and computational resources, posing challenges for deployment on battery-powered devices due to high energy consumption. Recent advancements have explored low bit-width data formats to simplify hardware implementation and reduce energy usage. However, accuracy loss with extremely low bit-widths, such as 1-bit, which enable basic operations implementation like multiplication with a single gate, remains a significant obstacle. Stochastic computing presents a promising alternative, allowing fundamental DNN operations like multiplication and accumulation to be performed with a single gate, leading to ultra-efficient computation. Despite its potential, this approach has not yet gained widespread attention and needs further investigation, especially for recent deep networks such as large language models.

 

Supervisor:

M. H. Sadi, C. Weis

 
Student:

Javier Ferrer Ortega

Year:

2024