Diverse Weight Averaging for Out-of-Distribution Generalization
This addresses the issue of poor generalization under distribution shifts for computer vision applications, representing an incremental improvement over existing weight averaging methods.
The paper tackles the problem of neural networks struggling with out-of-distribution generalization in computer vision by proposing Diverse Weight Averaging (DiWA), a weight averaging strategy that increases functional diversity across models, and it consistently improves state-of-the-art performance on the DomainBed benchmark without inference overhead.
Standard neural networks struggle to generalize under distribution shifts in computer vision. Fortunately, combining multiple networks can consistently improve out-of-distribution generalization. In particular, weight averaging (WA) strategies were shown to perform best on the competitive DomainBed benchmark; they directly average the weights of multiple networks despite their nonlinearities. In this paper, we propose Diverse Weight Averaging (DiWA), a new WA strategy whose main motivation is to increase the functional diversity across averaged models. To this end, DiWA averages weights obtained from several independent training runs: indeed, models obtained from different runs are more diverse than those collected along a single run thanks to differences in hyperparameters and training procedures. We motivate the need for diversity by a new bias-variance-covariance-locality decomposition of the expected error, exploiting similarities between WA and standard functional ensembling. Moreover, this decomposition highlights that WA succeeds when the variance term dominates, which we show occurs when the marginal distribution changes at test time. Experimentally, DiWA consistently improves the state of the art on DomainBed without inference overhead.