AdaFlood: Adaptive Flood Regularization
This work addresses the limitation of one-size-fits-all flood regularization for neural network training, offering a more nuanced approach that could benefit practitioners in various domains, though it is incremental over existing flood methods.
The paper tackles the problem of using a constant flood level for all training samples in flood regularization, which assumes equal sample difficulty, by introducing AdaFlood, an adaptive method that sets flood levels per sample based on difficulty. Experiments across four input modalities show its versatility in improving generalization.
Although neural networks are conventionally optimized towards zero training loss, it has been recently learned that targeting a non-zero training loss threshold, referred to as a flood level, often enables better test time generalization. Current approaches, however, apply the same constant flood level to all training samples, which inherently assumes all the samples have the same difficulty. We present AdaFlood, a novel flood regularization method that adapts the flood level of each training sample according to the difficulty of the sample. Intuitively, since training samples are not equal in difficulty, the target training loss should be conditioned on the instance. Experiments on datasets covering four diverse input modalities - text, images, asynchronous event sequences, and tabular - demonstrate the versatility of AdaFlood across data domains and noise levels.