Distributionally Robust Recurrent Decoders with Random Network Distillation
This addresses the practical problem of distribution shift in language models, which is incremental as it builds on existing OOD detection techniques.
The paper tackles the problem of neural language models degrading under distribution shift by proposing a method that uses OOD detection with Random Network Distillation to automatically disregard OOD context during inference, demonstrating improvements on multiple language modeling datasets.
Neural machine learning models can successfully model language that is similar to their training distribution, but they are highly susceptible to degradation under distribution shift, which occurs in many practical applications when processing out-of-domain (OOD) text. This has been attributed to "shortcut learning": relying on weak correlations over arbitrary large contexts. We propose a method based on OOD detection with Random Network Distillation to allow an autoregressive language model to automatically disregard OOD context during inference, smoothly transitioning towards a less expressive but more robust model as the data becomes more OOD while retaining its full context capability when operating in-distribution. We apply our method to a GRU architecture, demonstrating improvements on multiple language modeling (LM) datasets.