Work

Convergence of Gradient Descent for Recurrent Neural Networks: A Nonasymptotic Analysis

Year: 2025
Type: article
Abstract: . We analyze recurrent neural networks with diagonal hidden-to-hidden weight matrices, trained with gradient descent in the supervised learning setting, and prove that gradient descent can achieve optimality without massive overparameterization. Our in-depth nonasymptotic analysis (i) provides improved bounds on the network size [Formula: see text] in terms of the sequence length [Formula: see text], sample size [Formula: see text], and ambient dimension [Formula: see text], and (ii) identifies the significant impact of long-term dependencies in the dynamical system on the convergence and network width bounds characterized by a cutoff point that depends on the Lipschitz continuity of the activation function. Remarkably, this analysis reveals that an appropriately initialized recurrent neural network trained with [Formula: see text] samples can achieve optimality with a network size [Formula: see text] that scales only logarithmically with [Formula: see text]. This sharply contrasts with the prior works that require high-order polynomial dependency of [Formula: see text] on [Formula: see text] to establish strong regularity conditions. Our results are based on an explicit characterization of the class of dynamical systems that can be approximated and learned by recurrent neural networks via norm-constrained transportation mappings, and establishing local smoothness properties of the hidden state with respect to the learnable parameters.
Language: English
Open Access status: green
OpenAlex ID: W4411607431
FWCI: 1.957
Cites: 36
Cited by: 14
Related to: 9