Path Space for Recurrent Neural Networks with ReLU Activations

It is well known that neural networks with rectified linear units (ReLU) activation functions are positively scale-invariant (i.e., the neural network is invariant to positive rescaling of weights). Optimization algorithms like stochastic gradient descent that optimize the neural networks in the vector space of weights, which are not positively scale-invariant. To solve this mismatch, a new parameter space called path space has been proposed for feedforward and convolutional neural networks. The path space is positively scale-invariant and optimization algorithms operating in path space have been shown to be superior than that in the original weight space. However, the theory of path space and the corresponding optimization algorithm cannot be naturally extended to more complex neural networks, like Recurrent Neural Networks(RNN) due to the recurrent structure and the parameter sharing scheme over time. In this work, we aim to construct path space for RNN with ReLU activations so that we can employ optimization algorithms in path space. To achieve the goal, we propose leveraging the reduction graph of RNN which removes the influence of time-steps, and prove that all the values of whose paths can serve as a sufficient representation of the RNN with ReLU activations. We then prove that the path space for RNN is composed by the basis paths in reduction graph, and design a \emph{Skeleton Method} to identify the basis paths efficiently. With the identified basis paths, we develop the optimization algorithm in path space for RNN models. Our experiments on several benchmark datasets show that we can obtain significantly more effective RNN models in this way than using optimization methods in the weight space.