The Future is Log-Gaussian: ResNets and Their Infinite-Depth-and-Width Limit at Initialization
This provides a theoretical foundation for analyzing deep ResNets, addressing a bottleneck in understanding modern architectures for researchers in deep learning theory.
The paper tackles the problem that the Gaussian process approximation for neural networks deteriorates with increasing depth, showing that ReLU ResNets in the infinite-depth-and-width limit exhibit log-Gaussian behavior at initialization, with Monte Carlo simulations confirming this captures standard ResNet properties better than the Gaussian limit. It reveals hypoactivation (fewer than half of ReLUs activated) and interlayer correlations that exponentially increase output variance, leading to the introduction of Balanced ResNets to eliminate these issues.
Theoretical results show that neural networks can be approximated by Gaussian processes in the infinite-width limit. However, for fully connected networks, it has been previously shown that for any fixed network width, $n$, the Gaussian approximation gets worse as the network depth, $d$, increases. Given that modern networks are deep, this raises the question of how well modern architectures, like ResNets, are captured by the infinite-width limit. To provide a better approximation, we study ReLU ResNets in the infinite-depth-and-width limit, where both depth and width tend to infinity as their ratio, $d/n$, remains constant. In contrast to the Gaussian infinite-width limit, we show theoretically that the network exhibits log-Gaussian behaviour at initialization in the infinite-depth-and-width limit, with parameters depending on the ratio $d/n$. Using Monte Carlo simulations, we demonstrate that even basic properties of standard ResNet architectures are poorly captured by the Gaussian limit, but remarkably well captured by our log-Gaussian limit. Moreover, our analysis reveals that ReLU ResNets at initialization are hypoactivated: fewer than half of the ReLUs are activated. Additionally, we calculate the interlayer correlations, which have the effect of exponentially increasing the variance of the network output. Based on our analysis, we introduce Balanced ResNets, a simple architecture modification, which eliminates hypoactivation and interlayer correlations and is more amenable to theoretical analysis.