PSO–GP: An evolutionary optimization-based activation function approximation for hardware efficient neural network

Mahendra Kumar Gurve

doi:10.1016/j.compeleceng.2026.111404

PSO–GP: An evolutionary optimization-based activation function approximation for hardware efficient neural network

2026-07-30

0

PRE

AI

DOI：10.1016/j.compeleceng.2026.111404

Original

Original request for help

Save

Abstract

En 中文

• A unified PSO–GP framework is proposed to jointly optimize interval partitioning and approximation coefficients, eliminating manual design and enabling global optimization. • The proposed design achieves significant hardware efficiency with up to 64.37% LookUp Table reduction and 57.1% energy savings, suitable for FPGA and ASIC implementations. • The method attains high approximation accuracy (Mean Absolute Error of 2.77×10−4 with 13 segments) while preserving neural network performance (up to 98.45% accuracy on MNIST).

PSO–GP: An evolutionary optimization-based activation function approximation for hardware efficient neural network

Abstract

Journal

Organization