Neural Architecture Search of Sample Reweighting Networks for Complex Distribution Shift

This study presents an enhancement to the Meta-Weight-Net (MW-Net) through neural architecture search to improve its performance under complex distribution shifts, specifically addressing label noise and class imbalance. By employing a tree-structured Parzen estimator, the research optimizes the number of hidden layers and nodes, as well as identifies the most effective intermediate layer for input into MW-Net. Experimental results on modified CIFAR-10 and CIFAR-100 datasets indicate that this approach significantly mitigates performance degradation, offering practitioners a robust method for sample reweighting in challenging scenarios.