Recent progress in robust statistical learning has mainly tackled convex problems, like mean estimation or linear regression, with non-convex challenges receiving less attention. Phase retrieval exemplifies such a non-convex problem, requiring the recovery of a signal from only the magnitudes of its linear measurements, without phase (sign) information. While several non-convex methods, especially those involving the Wirtinger Flow algorithm, have been proposed for noiseless or mild noise settings, developing solutions for heavy-tailed noise and adversarial corruption remains an open challenge. In this paper, we investigate an approach that leverages robust gradient descent techniques to improve the Wirtinger Flow algorithm's ability to simultaneously cope with fourth moment bounded noise and adversarial contamination in both the inputs (covariates) and outputs (responses). We address two scenarios: known zero-mean noise and completely unknown noise. For the latter, we propose a preprocessing step that alters the problem into a new format that does not fit traditional phase retrieval approaches but can still be resolved with a tailored version of the algorithm for the zero-mean noise context.

We tackle the general differentiable meta learning problem that is ubiquitous in modern deep learning, including hyperparameter optimization, loss function learning, few-shot learning, invariance learning and more. These problems are often formalized as Bi-Level optimizations (BLO). We introduce a novel perspective by turning a given BLO problem into a stochastic optimization, where the inner loss function becomes a smooth probability distribution, and the outer loss becomes an expected loss over the inner distribution. To solve this stochastic optimization, we adopt Stochastic Gradient Langevin Dynamics (SGLD) MCMC to sample inner distribution, and propose a recurrent algorithm to compute the MC-estimated hypergradient. Our derivation is similar to forward-mode differentiation, but we introduce a new first-order approximation that makes it feasible for large models without needing to store huge Jacobian matrices. The main benefits are two-fold: i) Our stochastic formulation takes into account uncertainty, which makes the method robust to suboptimal inner optimization or non-unique multiple inner minima due to overparametrization; ii) Compared to existing methods that often exhibit unstable behavior and hyperparameter sensitivity in practice, our method leads to considerably more reliable solutions. We demonstrate that the new approach achieves promising results on diverse meta learning problems and easily scales to learning 87M hyperparameters in the case of Vision Transformers.

Sharpness-Aware Minimization (SAM) has substantially improved the generalization of neural networks under various settings. Despite the success, its effectiveness remains poorly understood. In this work, we discover an intriguing phenomenon in the training dynamics of SAM, shedding lights on understanding its implicit bias towards flatter minima over Stochastic Gradient Descent (SGD). We conjecture that the optimization method chosen in the late phase is more crucial in shaping the final solution's properties. Based on this viewpoint, we extend our findings from SAM to Adversarial Training. We provide source code in supplementary materials and will release checkpoints in future. Recently, it has been observed that the generalization of neural networks is closely tied to the sharpness of the loss landscape (Keskar et al., 2017; Zhang et al., 2017; Neyshabur et al., 2017; Jiang et al., 2020). This has led to the development of many gradient-based optimization algorithms that explicitly/implicitly regularize the sharpness of solutions. In particular, Foret et al. (2021) proposed Sharpness-Aware Minimization (SAM), which has substantially improved the generalization and robustness (Zhang et al., 2024) of neural networks across many tasks, including computer vision (Foret et al., 2021; Chen et al., 2022; Kaddour et al., 2022) and natural language processing (Bahri et al., 2022). Despite the empirical success of SAM, its effectiveness is not yet fully understood. Andriushchenko & Flammarion (2022) has shown that existing theoretical justifications based on PAC-Bayes generalization bounds (Foret et al., 2021; Wu et al., 2020a) are incomplete in explaining the superior performance of SAM.

In this work, we investigate a particular implicit bias in the gradient descent training process, which we term "Feature Averaging", and argue that it is one of the principal factors contributing to non-robustness of deep neural networks. Despite the existence of multiple discriminative features capable of classifying data, neural networks trained by gradient descent exhibit a tendency to learn the average (or certain combination) of these features, rather than distinguishing and leveraging each feature individually. In particular, we provide a detailed theoretical analysis of the training dynamics of gradient descent in a two-layer ReLU network for a binary classification task, where the data distribution consists of multiple clusters with orthogonal cluster center vectors. We rigorously prove that gradient descent converges to the regime of feature averaging, wherein the weights associated with each hidden-layer neuron represent an average of the cluster centers (each center corresponding to a distinct feature). It leads the network classifier to be non-robust due to an attack that aligns with the negative direction of the averaged features. Furthermore, we prove that, with the provision of more granular supervised information, a two-layer multi-class neural network is capable of learning individual features, from which one can derive a binary classifier with the optimal robustness under our setting. Besides, we also conduct extensive experiments using synthetic datasets, MNIST and CIFAR-10 to substantiate the phenomenon of feature averaging and its role in adversarial robustness of neural networks. We hope the theoretical and empirical insights can provide a deeper understanding of the impact of the gradient descent training on feature learning process, which in turn influences the robustness of the network, and how more detailed supervision may enhance model robustness.

Fine-tuning is powerful for adapting large language models to downstream tasks, but it often results in huge memory usages. A promising approach to mitigate this is using Zeroth-Order (ZO) optimization, which estimates gradients to replace First-Order (FO) gradient calculations, albeit with longer training time due to its stochastic nature. By revisiting the Memory-efficient ZO (MeZO) optimizer, we discover that the full-parameter perturbation and updating processes consume over 50% of its overall fine-tuning time cost. Based on these observations, we introduce a novel layer-wise sparse computation and memory efficient ZO optimizer, named LeZO. LeZO treats layers as fundamental units for sparsification and dynamically perturbs different parameter subsets in each step to achieve full-parameter fine-tuning. LeZO incorporates layer-wise parameter sparsity in the process of simultaneous perturbation stochastic approximation (SPSA) and ZO stochastic gradient descent (ZO-SGD). It achieves accelerated computation during perturbation and updating processes without additional memory overhead. We conduct extensive experiments with the OPT model family on the SuperGLUE benchmark and two generative tasks. The experiments show that LeZO accelerates training without compromising the performance of ZO optimization. Specifically, it achieves over 3x speedup compared to MeZO on the SST-2, BoolQ, and Copa tasks.

The sharpest known high probability excess risk bounds are up to $O\left( 1/n \right)$ for empirical risk minimization and projected gradient descent via algorithmic stability (Klochkov \& Zhivotovskiy, 2021). In this paper, we show that high probability excess risk bounds of order up to $O\left( 1/n^2 \right)$ are possible. We discuss how high probability excess risk bounds reach $O\left( 1/n^2 \right)$ under strongly convexity, smoothness and Lipschitz continuity assumptions for empirical risk minimization, projected gradient descent and stochastic gradient descent. Besides, to the best of our knowledge, our high probability results on the generalization gap measured by gradients for nonconvex problems are also the sharpest.

The information exponent (Ben Arous et al. [2021]) -- which is equivalent to the lowest degree in the Hermite expansion of the link function for Gaussian single-index models -- has played an important role in predicting the sample complexity of online stochastic gradient descent (SGD) in various learning tasks. In this work, we demonstrate that, for multi-index models, focusing solely on the lowest degree can miss key structural details of the model and result in suboptimal rates. Specifically, we consider the task of learning target functions of form $f_*(\mathbf{x}) = \sum_{k=1}^{P} \phi(\mathbf{v}_k^* \cdot \mathbf{x})$, where $P \ll d$, the ground-truth directions $\{ \mathbf{v}_k^* \}_{k=1}^P$ are orthonormal, and only the second and $2L$-th Hermite coefficients of the link function $\phi$ can be nonzero. Based on the theory of information exponent, when the lowest degree is $2L$, recovering the directions requires $d^{2L-1}\mathrm{poly}(P)$ samples, and when the lowest degree is $2$, only the relevant subspace (not the exact directions) can be recovered due to the rotational invariance of the second-order terms. In contrast, we show that by considering both second- and higher-order terms, we can first learn the relevant space via the second-order terms, and then the exact directions using the higher-order terms, and the overall sample and complexity of online SGD is $d \mathrm{poly}(P)$.

Matrix-valued optimization tasks, including those involving symmetric positive definite (SPD) matrices, arise in a wide range of applications in machine learning, data science and statistics. Classically, such problems are solved via constrained Euclidean optimization, where the domain is viewed as a Euclidean space and the structure of the matrices (e.g., positive definiteness) enters as constraints. More recently, geometric approaches that leverage parametrizations of the problem as unconstrained tasks on the corresponding matrix manifold have been proposed. While they exhibit algorithmic benefits in many settings, they cannot directly handle additional constraints, such as inequality or sparsity constraints. A remedy comes in the form of constrained Riemannian optimization methods, notably, Riemannian Frank-Wolfe and Projected Gradient Descent. However, both algorithms require potentially expensive subroutines that can introduce computational bottlenecks in practise. To mitigate these shortcomings, we introduce a class of structured regularizers, based on symmetric gauge functions, which allow for solving constrained optimization on the SPD manifold with faster unconstrained methods. We show that our structured regularizers can be chosen to preserve or induce desirable structure, in particular convexity and "difference of convex" structure. We demonstrate the effectiveness of our approach in numerical experiments.

We introduce a method called TracIn that computes the influence of a training example on a prediction made by the model. The idea is to trace how the loss on the test point changes during the training process whenever the training example of interest was utilized. We provide a scalable implementation of TracIn via: (a) a first-order gradient approximation to the exact computation, (b) saved checkpoints of standard training procedures, and (c) cherry-picking layers of a deep neural network. In contrast with previously proposed methods, TracIn is simple to implement; all it needs is the ability to work with gradients, checkpoints, and loss functions. It applies to any machine learning model trained using stochastic gradient descent or a variant of it, agnostic of architecture, domain and task.

We investigate the spectral properties of linear-width feed-forward neural networks, where the sample size is asymptotically proportional to network width. Empirically, we show that the spectra of weight in this high dimensional regime are invariant when trained by gradient descent for small constant learning rates; we provide a theoretical justification for this observation and prove the invariance of the bulk spectra for both conjugate and neural tangent kernels. We demonstrate similar characteristics when training with stochastic gradient descent with small learning rates. When the learning rate is large, we exhibit the emergence of an outlier whose corresponding eigenvector is aligned with the training data structure. We also show that after adaptive gradient training, where a lower test error and feature learning emerge, both weight and kernel matrices exhibit heavy tail behavior.