Evolutionary Neural Cascade Search across Supernetworks

TL;DR

ENCAS introduces an evolutionary search method for neural network cascades that optimizes accuracy and computational efficiency across multiple supernetworks, outperforming previous NAS approaches on standard benchmarks.

Contribution

It extends Neural Architecture Transfer by enabling multi-supernetwork cascade search to improve accuracy and efficiency trade-offs in neural network design.

Findings

Achieves Pareto dominance over state-of-the-art NAS models up to 1.5 GFLOPs.

Increases maximum accuracy from 88.6% to 89.0% on ImageNet.

Reduces computation effort by 18% from 362 to 296 GFLOPs.

Abstract

To achieve excellent performance with modern neural networks, having the right network architecture is important. Neural Architecture Search (NAS) concerns the automatic discovery of task-specific network architectures. Modern NAS approaches leverage supernetworks whose subnetworks encode candidate neural network architectures. These subnetworks can be trained simultaneously, removing the need to train each network from scratch, thereby increasing the efficiency of NAS. A recent method called Neural Architecture Transfer (NAT) further improves the efficiency of NAS for computer vision tasks by using a multi-objective evolutionary algorithm to find high-quality subnetworks of a supernetwork pretrained on ImageNet. Building upon NAT, we introduce ENCAS - Evolutionary Neural Cascade Search. ENCAS can be used to search over multiple pretrained supernetworks to achieve a trade-off front of cascades of different neural network architectures, maximizing accuracy while minimizing FLOPs count. We test ENCAS on common computer vision benchmarks (CIFAR-10, CIFAR-100, ImageNet) and achieve Pareto dominance over previous state-of-the-art NAS models up to 1.5 GFLOPs. Additionally, applying ENCAS to a pool of 518 publicly available ImageNet classifiers leads to Pareto dominance in all computation regimes and to increasing the maximum accuracy from 88.6% to 89.0%, accompanied by an 18\% decrease in computation effort from 362 to 296 GFLOPs. Our code is available at https://github.com/AwesomeLemon/ENCAS