Does Interference Exist When Training a Once-For-All Network?

TL;DR

This paper reevaluates the impact of interference during training of Once-For-All networks, finding that architecture bias is more critical than interference mitigation, and introduces a simple sampling method that improves efficiency.

Contribution

It challenges the belief that interference mitigation is crucial in OFA training and proposes Random Subnet Sampling, which simplifies training and enhances performance.

Findings

Interference mitigation strategies have limited impact on subnet performance.

Random Subnet Sampling outperforms Progressive Shrinking in several datasets.

RSS significantly reduces training time with minimal performance loss.

Abstract

The Once-For-All (OFA) method offers an excellent pathway to deploy a trained neural network model into multiple target platforms by utilising the supernet-subnet architecture. Once trained, a subnet can be derived from the supernet (both architecture and trained weights) and deployed directly to the target platform with little to no retraining or fine-tuning. To train the subnet population, OFA uses a novel training method called Progressive Shrinking (PS) which is designed to limit the negative impact of interference during training. It is believed that higher interference during training results in lower subnet population accuracies. In this work we take a second look at this interference effect. Surprisingly, we find that interference mitigation strategies do not have a large impact on the overall subnet population performance. Instead, we find the subnet architecture selection bias during training to be a more important aspect. To show this, we propose a simple-yet-effective method called Random Subnet Sampling (RSS), which does not have mitigation on the interference effect. Despite no mitigation, RSS is able to produce a better performing subnet population than PS in four small-to-medium-sized datasets; suggesting that the interference effect does not play a pivotal role in these datasets. Due to its simplicity, RSS provides a $1.9\times$ reduction in training times compared to PS. A $6.1\times$ reduction can also be achieved with a reasonable drop in performance when the number of RSS training epochs are reduced. Code available at https://github.com/Jordan-HS/RSS-Interference-CVPRW2022.