Approximate computing for machine learning

Abstract

Approximate Computing has emerged as a promising paradigm to enhance computational efficiency by introducing controlled inaccuracies in arithmetic operations. This study explores the application of static approximate adders (AAs) within a machine learning context, specifically the K-means clustering algorithm, to assess the trade-offs between clustering accuracy and computational resource savings. A total of 17 AAs, including two newly proposed designs (BPAA-LSP1 and NAA), along with a conventional accurate adder, were integrated into a K-means clustering algorithm and evaluated across four benchmark datasets. Both software simulations and hardware implementations were conducted, measuring key performance metrics such as within-cluster sum of squares (WCSS), power, delay, and area. Results demonstrate that approximate adders maintain strong clustering performance while significantly reducing power consumption (up to 60% lower power-delay product) and area usage (up to 50% lower area-delay product). Notably, the newly proposed BPAA-LSP1 and NAA adders achieve an outstanding balance between clustering accuracy and computational efficiency, comparable to those achieved using the accurate adder. BPAA-LSP1 demonstrates reductions of 22.2% in power, 21.6% in area, and 26.3% in delay, while NAA achieves even greater efficiency, with 31.0% lower power consumption, 21.6% reduced area, and a 37.1% decrease in delay. These results suggest that specific characteristics of the error generated by these AAs may be beneficial or have error-compensatory effects for iterative machine learning tasks. These findings indicate that approximate adders, particularly BPAA-LSP1 and NAA, could serve as viable alternatives in energy-efficient, error-tolerant machine learning applications.