Batch Normalization (BN) is a cornerstone of modern deep learning, essential for training architectures ranging from ResNets to Transformers. Yet, despite its ubiquity, its theoretical underpinnings have remained surprisingly elusive. In this paper, researchers develop a precise Mean Field Theory (MFT) to characterize signal and gradient propagation in wide, batch-normalized networks at initialization. Counterintuitively, they reveal that BN actually causes gradients to explode exponentially with depth, making deep networks without skip connections untrainable at initialization—a crisis averted only by a rapid dynamical equilibrium reached after a single optimization step.
The Challenge of Analyzing Batch Normalization
Mean Field Theory is a standard tool for studying neural networks in the "infinite width" limit. By treating weights as random variables, researchers can predict how activation distributions evolve as they pass through dozens of layers. While MFT has successfully modeled vanilla feed-forward networks and ResNets, Batch Normalization introduces a significant mathematical hurdle: it normalizes across an entire batch of data rather than a single input. This introduces non-local dependencies between inputs, breaking standard MFT assumptions. This paper bridges that gap, providing a formal framework to analyze BN's true impact on trainability.
Theoretical Framework
To handle the complex batch-wise dependencies of BN, the authors model network pre-activations as jointly Gaussian random variables. Solving the resulting integrals requires a sophisticated mathematical toolkit, including Laplace, Fourier, and Gegenbauer transforms. Gegenbauer polynomials are particularly suited for this task because BN inherently projects data onto a high-dimensional sphere, and these polynomials are the natural language for functions on spheres.
Furthermore, the authors introduce a new class of linear operators called "ultrasymmetric operators" to describe the evolution of covariance between different batch samples. This allows them to identify fixed points—states where the internal statistics of the network stabilize.
The Gradient Explosion Paradox
The original BN paper hypothesized that normalization aids training by keeping layer Jacobians near 1. This paper mathematically proves the opposite: at initialization, BN forces the singular values of layer Jacobians away from 1, leading to exponential gradient explosion.
As data propagates deeper into a BN network, different samples in a batch become increasingly decorrelated. The network essentially "forgets" the relationships between inputs, entering a chaotic state. The authors calculate the exact rate of this growth: for a ReLU network with a large batch size, the gradient norm grows by a factor of approximately $1.467$ per layer. Consequently, vanilla BN networks (without skip connections) become untrainable beyond roughly 50 layers due to numerical instability.
The "First Step" Savior and the Linear Cure
If gradients explode exponentially, how do BN networks train at all? The authors discovered a fascinating phenomenon: after just one step of gradient descent, the network weights adjust enough to reach a stable dynamical equilibrium. Gradients drop from astronomical values to a manageable range, allowing training to proceed.
Additionally, the authors found that gradient explosion is exacerbated by highly non-linear activation functions. By tuning the network closer to the linear regime at initialization (e.g., reducing pre-activation scales in a tanh network), the explosion rate decreases, enabling the training of deeper networks without residual connections.
Empirical Validation
The theoretical predictions were rigorously validated using Monte Carlo simulations on randomly initialized networks (width 1000) to confirm signal propagation laws. Experiments on MNIST and CIFAR-10 confirmed the predicted "maximum trainable depth" limits, showing exceptional agreement between actual test accuracy and the theoretical depth scale derived from the Mean Field equations.
Limitations and Future Directions
While this work provides a rigorous analysis of network initialization, it primarily focuses on the state before training begins. The discovery of the "first step" dynamical equilibrium is profound, but a comprehensive theory of learning dynamics throughout the entire training process remains an open challenge.
Furthermore, the analysis assumes infinite width; exploring finite-width effects in smaller networks is a necessary next step. Ultimately, this foundational work not only clarifies the mechanics of Batch Normalization but also paves the way for advanced techniques, such as exact Bayesian inference in wide batch-normalized networks.
Summary updated