Meta-Heuristic Optimization of CNNs with Approximate Error Distributed Multipliers

Convolutional Neural Networks (CNNs) offer un-matched feature extraction capability, making them predomi-nantly used in several applications involving time-series signals. However, CNN hardware implementation remains challenging as it involves extensive computations. Among them, Multipliers are the most power-consuming and high compute-latency units. Adopting approximate multipliers (AM) relaxes hardware de-mands but suffers from the drop in accuracy. Therefore, choosing AMs becomes pivotal. Leveraging a precise combination of AMs along the convolutional layers instead of uniform multipliers throughout the network enhances network performance. In this work, seven meta-heuristic algorithms, including genetic and non-genetic, are explored to deduce optimal solutions and are benchmarked with exhaustive design-space generated solutions. Single-objective optimization algorithms, including Genetic Algorithm (GA), Differential Evolution (DE), and Particle Swarm Optimization (PSO), also multi-objective algorithms, including Non-dominated Sorting Genetic Algorithm-II (NSGA-II), Ref-erence point-based NSGA-II (RNSGA-II), Adaptive Geometry Estimation based Multi-Objective Evolutionary Algorithm (AGE-MOEA), and AGE-MOEA-II, are exploited to identify pareto-optimal solutions comprising of AM sequences that balance hardware parameters and CNN accuracy. All hardware CNN designs generated for objectives, including hardware parameters and model accuracy, targeted for three CNN models are discussed in this work. The proposed meta-heuristic approach presents a unique framework for designing hardware-efficient and error-resilient AI-on-chip design.