Deep ReLU Networks Have Surprisingly Few Activation Patterns
This work addresses the gap between theoretical and practical expressivity in deep learning for researchers, indicating that current methods may be inherently limited.
The paper shows that the average number of activation patterns in ReLU networks at initialization is bounded by the total number of neurons raised to the input dimension, which is independent of depth and empirically tight during training, suggesting that realizing the full theoretical expressivity of deep networks may not be feasible in practice.
The success of deep networks has been attributed in part to their expressivity: per parameter, deep networks can approximate a richer class of functions than shallow networks. In ReLU networks, the number of activation patterns is one measure of expressivity; and the maximum number of patterns grows exponentially with the depth. However, recent work has showed that the practical expressivity of deep networks - the functions they can learn rather than express - is often far from the theoretical maximum. In this paper, we show that the average number of activation patterns for ReLU networks at initialization is bounded by the total number of neurons raised to the input dimension. We show empirically that this bound, which is independent of the depth, is tight both at initialization and during training, even on memorization tasks that should maximize the number of activation patterns. Our work suggests that realizing the full expressivity of deep networks may not be possible in practice, at least with current methods.