Approximate implementations of neural networks

Abstract

This research explores the application of approximate computing in neural networks, focusing on both classical models and the innovative Truth Table Nets (TTnet). The study aims to evaluate how approximation techniques can optimize computational efficiency without compromising the accuracy, a crucial balance due to rising demands of AI and ML applications. The research involved testing multiple approximations tools, revealing significant challenges from outdated software dependencies. As an alternate approach, custom Python programs were developed to generate and evaluate truth tables by modifying Boolean expressions through term and variable reduction. However, reproducing the original accuracy of TTnet with the approximated TT-rules and associated weights proved difficult, limiting the project’s ability to assess the full impact of these optimizations. Despite these setbacks, the project offers insights into the challenges and potential of approximate computing in novel neural network architectures, paving the way for future exploration in the domain.