Kuo Gai

Hu Tan, Kuo Gai, Shihua Zhang

Grokking suggests that fitting the training data and learning a simple underlying rule may occur on different time scales. We formalize this phenomenon by separating the fast decay of the classification loss from the slower simplification of the learned representation, and we call the resulting pair of stopping times two training clocks. For deep linear networks, we show that a post-margin gap-growth or one-step tail-contraction condition reduces the cross-entropy loss to level epsilon on a logarithmic time scale. In contrast, when layerwise weight decay is present, the induced regularization on the end-to-end map can be expressed as a Schatten-type penalty; under a sharp late-time Kurdyka-Lojasiewicz tail, this structural energy closes on a polynomial time scale. The two clocks, therefore, separate fitting from representation simplification. We then explain how the same mechanism can appear in ReLU MLPs. In regions where the activation patterns on the training set remain fixed, the network reduces to a linear model in the active coordinates. In a two-layer ReLU embedding model, chain-rule estimates further show that the classifier head can receive larger effective gradients than the embedding block under controlled downstream norms. This supports a two-stage mechanism in which the classifier fits first, while the representation continues to simplify later. We use modular addition as the main experimental setting. The deep linear theory provides the rigorous core of the analysis. But the ReLU results are formulated as conditional reductions that account for empirical behavior without claiming a global proof for nonlinear training dynamics.

Kuo Gai, Sicong Wang, Shihua Zhang

Deep neural networks (DNNs) are vulnerable to small adversarial perturbations of the inputs, posing a significant challenge to their reliability and robustness. Empirical methods such as adversarial training can defend against particular attacks but remain vulnerable to more powerful attacks. Alternatively, Lipschitz networks provide certified robustness to unseen perturbations but lack sufficient expressive power. To harness the advantages of both approaches, we design a novel two-step Optimal Transport induced Adversarial Defense (OTAD) model that can fit the training data accurately while preserving the local Lipschitz continuity. First, we train a DNN with a regularizer derived from optimal transport theory, yielding a discrete optimal transport map linking data to its features. By leveraging the map's inherent regularity, we interpolate the map by solving the convex integration problem (CIP) to guarantee the local Lipschitz property. OTAD is extensible to diverse architectures of ResNet and Transformer, making it suitable for complex data. For efficient computation, the CIP can be solved through training neural networks. OTAD opens a novel avenue for developing reliable and secure deep learning systems through the regularity of optimal transport maps. Empirical results demonstrate that OTAD can outperform other robust models on diverse datasets.

Kuo Gai, Shihua Zhang

In practice, deeper networks tend to be more powerful than shallow ones, but this has not been understood theoretically. In this paper, we find the analytical solution of a three-layer network with a matrix exponential activation function, i.e., $$ f(X)=W_3\exp(W_2\exp(W_1X)), X\in \mathbb{C}^{d\times d} $$ have analytical solutions for the equations $$ Y_1=f(X_1),Y_2=f(X_2) $$ for $X_1,X_2,Y_1,Y_2$ with only invertible assumptions. Our proof shows the power of depth and the use of a non-linear activation function, since one layer network can only solve one equation,i.e.,$Y=WX$.

Xiayang Li, Kuo Gai, Shihua Zhang

Shortcut learning causes deep learning models to rely on non-essential features within the data. However, its formation in deep neural network training still lacks theoretical understanding. In this paper, we provide a formal definition of core and shortcut features and employ evolutionary game theory to analyze the origins of shortcut bias by modeling data samples as players and their corresponding neural tangent features as strategies, assuming the existence of core and shortcut subnetworks. We find that gradient descent (GD) and stochastic gradient descent (SGD) lead to two distinct stochastically stable states, each corresponding to a different strategy. The former primarily optimizes the shortcut subnetwork, while the latter primarily optimizes the core subnetwork. We investigate the influence of these strategies on shortcut bias through a continuous stochastic differential equation, and reveal the impact of data noise and optimization noise on the formation of shortcut bias. In brief, our work employs evolutionary game theory to characterize the dynamics of shortcut bias formation and provides a theoretical view on its mitigation.

Sicong Wang, Kuo Gai, Shihua Zhang

Neural collapse (NC) is a simple and symmetric phenomenon for deep neural networks (DNNs) at the terminal phase of training, where the last-layer features collapse to their class means and form a simplex equiangular tight frame aligning with the classifier vectors. However, the relationship of the last-layer features to the data and intermediate layers during training remains unexplored. To this end, we characterize the geometry of intermediate layers of ResNet and propose a novel conjecture, progressive feedforward collapse (PFC), claiming the degree of collapse increases during the forward propagation of DNNs. We derive a transparent model for the well-trained ResNet according to that ResNet with weight decay approximates the geodesic curve in Wasserstein space at the terminal phase. The metrics of PFC indeed monotonically decrease across depth on various datasets. We propose a new surrogate model, multilayer unconstrained feature model (MUFM), connecting intermediate layers by an optimal transport regularizer. The optimal solution of MUFM is inconsistent with NC but is more concentrated relative to the input data. Overall, this study extends NC to PFC to model the collapse phenomenon of intermediate layers and its dependence on the input data, shedding light on the theoretical understanding of ResNet in classification problems.

Tianyu Ruan, Kuo Gai, Shihua Zhang

Why do deep networks generalize well? In contrast to classical generalization theory, we approach this fundamental question by examining not only inputs and outputs, but the evolution of internal features. Our study suggests a phenomenon of temporal consistency that predictions remain stable when shallow features from earlier checkpoints combine with deeper features from later ones. This stability is not a trivial convergence artifact. It acts as a form of implicit, structured augmentation that supports generalization. We show that temporal consistency extends to unseen and corrupted data, but collapses when semantic structure is destroyed (e.g., random labels). Statistical tests further reveal that SGD injects anisotropic noise aligned with a few principal directions, reinforcing its role as a source of structured variability. Together, these findings suggest a conceptual perspective that links feature dynamics to generalization, pointing toward future work on practical surrogates for measuring temporal feature evolution.

Kuo Gai, Shihua Zhang

Recent studies revealed the mathematical connection of deep neural network (DNN) and dynamic system. However, the fundamental principle of DNN has not been fully characterized with dynamic system in terms of optimization and generalization. To this end, we build the connection of DNN and continuity equation where the measure is conserved to model the forward propagation process of DNN which has not been addressed before. DNN learns the transformation of the input distribution to the output one. However, in the measure space, there are infinite curves connecting two distributions. Which one can lead to good optimization and generaliztion for DNN? By diving the optimal transport theory, we find DNN with weight decay attempts to learn the geodesic curve in the Wasserstein space, which is induced by the optimal transport map. Compared with plain network, ResNet is a better approximation to the geodesic curve, which explains why ResNet can be optimized and generalize better. Numerical experiments show that the data tracks of both plain network and ResNet tend to be line-shape in term of line-shape score (LSS), and the map learned by ResNet is closer to the optimal transport map in term of optimal transport score (OTS). In a word, we conclude a mathematical principle of deep learning is to learn the geodesic curve in the Wasserstein space; and deep learning is a great engineering realization of continuous transformation in high-dimensional space.

Kuo Gai, Shihua Zhang

Non-adversarial generative models such as variational auto-encoder (VAE), Wasserstein auto-encoders with maximum mean discrepancy (WAE-MMD), sliced-Wasserstein auto-encoder (SWAE) are relatively easy to train and have less mode collapse compared to Wasserstein auto-encoder with generative adversarial network (WAE-GAN). However, they are not very accurate in approximating the target distribution in the latent space because they don't have a discriminator to detect the minor difference between real and fake. To this end, we develop a novel non-adversarial framework called Tessellated Wasserstein Auto-encoders (TWAE) to tessellate the support of the target distribution into a given number of regions by the centroidal Voronoi tessellation (CVT) technique and design batches of data according to the tessellation instead of random shuffling for accurate computation of discrepancy. Theoretically, we demonstrate that the error of estimate to the discrepancy decreases when the numbers of samples $n$ and regions $m$ of the tessellation become larger with rates of $\mathcal{O}(\frac{1}{\sqrt{n}})$ and $\mathcal{O}(\frac{1}{\sqrt{m}})$, respectively. Given fixed $n$ and $m$, a necessary condition for the upper bound of measurement error to be minimized is that the tessellation is the one determined by CVT. TWAE is very flexible to different non-adversarial metrics and can substantially enhance their generative performance in terms of Fréchet inception distance (FID) compared to VAE, WAE-MMD, SWAE. Moreover, numerical results indeed demonstrate that TWAE is competitive to the adversarial model WAE-GAN, demonstrating its powerful generative ability.

Chihao Zhang, Kuo Gai, Shihua Zhang

Principal component analysis (PCA) is one of the most widely used dimension reduction and multivariate statistical techniques. From a probabilistic perspective, PCA seeks a low-dimensional representation of data in the presence of independent identical Gaussian noise. Probabilistic PCA (PPCA) and its variants have been extensively studied for decades. Most of them assume the underlying noise follows a certain independent identical distribution. However, the noise in the real world is usually complicated and structured. To address this challenge, some variants of PCA for data with non-IID noise have been proposed. However, most of the existing methods only assume that the noise is correlated in the feature space while there may exist two-way structured noise. To this end, we propose a powerful and intuitive PCA method (MN-PCA) through modeling the graphical noise by the matrix normal distribution, which enables us to explore the structure of noise in both the feature space and the sample space. MN-PCA obtains a low-rank representation of data and the structure of noise simultaneously. And it can be explained as approximating data over the generalized Mahalanobis distance. We develop two algorithms to solve this model: one maximizes the regularized likelihood, the other exploits the Wasserstein distance, which is more robust. Extensive experiments on various data demonstrate their effectiveness.