Trainability of ReLU networks and Data-dependent Initialization
This addresses the challenge of training deep neural networks for researchers and practitioners, providing theoretical insights into initialization and over-parameterization, though it is incremental as it builds on existing work in network initialization.
The paper tackles the problem of trainability in ReLU networks by introducing concepts of neuron death and showing that over-parameterization is necessary and sufficient for minimizing training loss, with numerical examples demonstrating effectiveness.
In this paper, we study the trainability of rectified linear unit (ReLU) networks. A ReLU neuron is said to be dead if it only outputs a constant for any input. Two death states of neurons are introduced; tentative and permanent death. A network is then said to be trainable if the number of permanently dead neurons is sufficiently small for a learning task. We refer to the probability of a network being trainable as trainability. We show that a network being trainable is a necessary condition for successful training and the trainability serves as an upper bound of successful training rates. In order to quantify the trainability, we study the probability distribution of the number of active neurons at the initialization. In many applications, over-specified or over-parameterized neural networks are successfully employed and shown to be trained effectively. With the notion of trainability, we show that over-parameterization is both a necessary and a sufficient condition for minimizing the training loss. Furthermore, we propose a data-dependent initialization method in the over-parameterized setting. Numerical examples are provided to demonstrate the effectiveness of the method and our theoretical findings.