On the stability properties of gated recurrent units neural networks (Q2059489): Difference between revisions

The paper provides sufficient sufficient conditions for input-to-state stability (ISS) and the incremental input-to-state stability (\(\delta\)ISS) of gated recurrent units (GRUs) neural networks. Both single-layer architecture \[ \begin{cases} x^+ = z \circ x + (1-z) \circ \phi(W_r u + U_r f \circ x + b_r)\\ z = \sigma(W_z u + U_z x + b_z)\\ f = \sigma(W_f u + U_f x + b_f)\\ y = U_0 x + b_0 \end{cases}, \] and its multi-layer counterpart are considered. Here \(x\in\mathbb R^{n_x}\), \(n_x \in \mathbb N\) is the state vector, \(u\in \mathbb R^{n_u}\), \(n_u \in \mathbb N\) is the input vector, \(y \in \mathbb R^{n_o}\), \(n_o \in \mathbb N\) is the output vector. The matrices \(W_*\), \(U_*\), and \(b_*\) are the weights and biases that parametrize the model. The derived sufficient stability conditions consist of nonlinear inequalities on the network's weights. They can be utilized in various ways, e.g., to verify stability of the trained network, or they can be enforced as constraints during the training procedure of a GRU. The resulting training procedure is tested on a quadruple tank nonlinear benchmark system.