Timezone: »

 
Poster
Noisy Recurrent Neural Networks
Soon Hoe Lim · N. Benjamin Erichson · Liam Hodgkinson · Michael Mahoney

Thu Dec 09 08:30 AM -- 10:00 AM (PST) @ Virtual
in Poster Session 6 »

We provide a general framework for studying recurrent neural networks (RNNs) trained by injecting noise into hidden states. Specifically, we consider RNNs that can be viewed as discretizations of stochastic differential equations driven by input data. This framework allows us to study the implicit regularization effect of general noise injection schemes by deriving an approximate explicit regularizer in the small noise regime. We find that, under reasonable assumptions, this implicit regularization promotes flatter minima; it biases towards models with more stable dynamics; and, in classification tasks, it favors models with larger classification margin. Sufficient conditions for global stability are obtained, highlighting the phenomenon of stochastic stabilization, where noise injection can improve stability during training. Our theory is supported by empirical results which demonstrate that the RNNs have improved robustness with respect to various input perturbations.

Keywords: Theory · Deep Learning · Machine Learning · Robustness

Author Information

Soon Hoe Lim (Nordita)
N. Benjamin Erichson (University of Pittsburgh)
Liam Hodgkinson (UC Berkeley)
Michael Mahoney (UC Berkeley)

More from the Same Authors