19
0

On the Role of Noise in the Sample Complexity of Learning Recurrent Neural Networks: Exponential Gaps for Long Sequences

Abstract

We consider the class of noisy multi-layered sigmoid recurrent neural networks with ww (unbounded) weights for classification of sequences of length TT, where independent noise distributed according to N(0,σ2)\mathcal{N}(0,\sigma^2) is added to the output of each neuron in the network. Our main result shows that the sample complexity of PAC learning this class can be bounded by O(wlog(T/σ))O (w\log(T/\sigma)). For the non-noisy version of the same class (i.e., σ=0\sigma=0), we prove a lower bound of Ω(wT)\Omega (wT) for the sample complexity. Our results indicate an exponential gap in the dependence of sample complexity on TT for noisy versus non-noisy networks. Moreover, given the mild logarithmic dependence of the upper bound on 1/σ1/\sigma, this gap still holds even for numerically negligible values of σ\sigma.

View on arXiv
Comments on this paper