The stochastic gradient descent (SGD) algorithm is the algorithm we use to train neural networks. However, it remains poorly understood how the SGD navigates the highly nonlinear and degenerate loss landscape of a neural network. In this work, we prove that the minibatch noise of SGD regularizes the solution towards a balanced solution whenever the loss function contains a rescaling symmetry. Because the difference between a simple diffusion process and SGD dynamics is the most significant when symmetries are present, our theory implies that the loss function symmetries constitute an essential probe of how SGD works. We then apply this result to derive the stationary distribution of stochastic gradient flow for a diagonal linear network with arbitrary depth and width. The stationary distribution exhibits complicated nonlinear phenomena such as phase transitions, broken ergodicity, and fluctuation inversion. These phenomena are shown to exist uniquely in deep networks, implying a fundamental difference between deep and shallow models.

This paper studies the stochastic optimization for decentralized nonconvex-strongly-concave (NC-SC) minimax problems over a multi-agent network. We propose an efficient algorithm, called the Decentralized Recursive-gradient descEnt Ascent Method (DREAM), which achieves the best-known theoretical guarantee for finding the $\epsilon$-stationary point of the primal function. The proposed method requires $\mathcal{O}(\min (\kappa^3\epsilon^{-3},\sqrt{N} \kappa^2 \epsilon^{-2} ))$ stochastic first-order oracle (SFO) calls and $\tilde{\mathcal{O}}(\kappa^2 \epsilon^{-2})$ communication rounds to find an $\epsilon$-stationary point, where $\kappa$ is the condition number. DREAM achieves the best-known complexity for both the online and offline setups.

Large skew-t factor copula models are attractive for the modeling of financial data because they allow for asymmetric and extreme tail dependence. We show that the copula implicit in the skew-t distribution of Azzalini and Capitanio (2003) allows for a higher level of pairwise asymmetric dependence than two popular alternative skew-t copulas. Estimation of this copula in high dimensions is challenging, and we propose a fast and accurate Bayesian variational inference (VI) approach to do so. The method uses a conditionally Gaussian generative representation of the skew-t distribution to define an augmented posterior that can be approximated accurately. A fast stochastic gradient ascent algorithm is used to solve the variational optimization. The new methodology is used to estimate copula models for intraday returns from 2017 to 2021 on 93 U.S. equities. The copula captures substantial heterogeneity in asymmetric dependence over equity pairs, in addition to the variability in pairwise correlations. We show that intraday predictive densities from the skew-t copula are more accurate than from some other copula models, while portfolio selection strategies based on the estimated pairwise tail dependencies improve performance relative to the benchmark index.

This paper studies Byzantine-robust stochastic optimization over a decentralized network, where every agent periodically communicates with its neighbors to exchange local models, and then updates its own local model by stochastic gradient descent (SGD). The performance of such a method is affected by an unknown number of Byzantine agents, which conduct adversarially during the optimization process. To the best of our knowledge, there is no existing work that simultaneously achieves a linear convergence speed and a small learning error. We observe that the learning error is largely dependent on the intrinsic stochastic gradient noise. Motivated by this observation, we introduce two variance reduction methods, stochastic average gradient algorithm (SAGA) and loopless stochastic variance-reduced gradient (LSVRG), to Byzantine-robust decentralized stochastic optimization for eliminating the negative effect of the stochastic gradient noise. The two resulting methods, BRAVO-SAGA and BRAVO-LSVRG, enjoy both linear convergence speeds and stochastic gradient noise-independent learning errors. Such learning errors are optimal for a class of methods based on total variation (TV)-norm regularization and stochastic subgradient update. We conduct extensive numerical experiments to demonstrate their effectiveness under various Byzantine attacks.

We explore the equivalence between neural networks and kernel methods by deriving the first exact representation of any finite-size parametric classification model trained with gradient descent as a kernel machine. We compare our exact representation to the well-known Neural Tangent Kernel (NTK) and discuss approximation error relative to the NTK and other non-exact path kernel formulations. We experimentally demonstrate that the kernel can be computed for realistic networks up to machine precision. We use this exact kernel to show that our theoretical contribution can provide useful insights into the predictions made by neural networks, particularly the way in which they generalize.

In supervised learning, the regularization path is sometimes used as a convenient theoretical proxy for the optimization path of gradient descent initialized from zero. In this paper, we study a modification of the regularization path for infinite-width 2-layer ReLU neural networks with nonzero initial distribution of the weights at different scales. By exploiting a link with unbalanced optimal-transport theory, we show that, despite the non-convexity of the 2-layer network training, this problem admits an infinite-dimensional convex counterpart. We formulate the corresponding functional-optimization problem and investigate its main properties. In particular, we show that, as the scale of the initialization ranges between $0$ and $+\infty$, the associated path interpolates continuously between the so-called kernel and rich regimes. Numerical experiments confirm that, in our setting, the scaling path and the final states of the optimization path behave similarly, even beyond these extreme points.

The ability to learn continuously from an incoming data stream without catastrophic forgetting is critical to designing intelligent systems. Many approaches to continual learning rely on stochastic gradient descent and its variants that employ global error updates, and hence need to adopt strategies such as memory buffers or replay to circumvent its stability, greed, and short-term memory limitations. To address this limitation, we have developed a biologically inspired lightweight neural network architecture that incorporates synaptic plasticity mechanisms and neuromodulation and hence learns through local error signals to enable online continual learning without stochastic gradient descent. Our approach leads to superior online continual learning performance on Split-MNIST, Split-CIFAR-10, and Split-CIFAR-100 datasets compared to other memory-constrained learning approaches and matches that of the state-of-the-art memory-intensive replay-based approaches. We further demonstrate the effectiveness of our approach by integrating key design concepts into other backpropagation-based continual learning algorithms, significantly improving their accuracy. Our results provide compelling evidence for the importance of incorporating biological principles into machine learning models and offer insights into how we can leverage them to design more efficient and robust systems for online continual learning. Online continual learning addresses the scenario where a system has to learn and process data that are continuously streamed, often without restrictions in terms of the distribution of data within and across tasks and without clearly identified task boundaries Mai et al. (2021); Chen et al. (2020); Aljundi et al. (2019a). Online continual learning algorithms seek to mitigate catastrophic forgetting at both the data-instance and task level Chen et al. (2020). In some cases, however, such as on-chip learning at the edge, additional considerations such as resource limitations in the hardware, data privacy, or data security are also important for online continual learning. A key challenge of online continual learning is that it runs counter to the optimal conditions required for optimization using stochastic gradient descent (SGD) Parisi et al. (2019), which struggles with non-stationary data streams Lindsey & Litwin-Kumar (2020). On the contrary, biological systems excel at online continual learning. Inspired by the structure and functionality of the mammal brain, several approaches have adopted replay strategies to counteract catastrophic forgetting during non-stationary tasks.

This paper presents a framework for smooth optimization of objectives with $\ell_q$ and $\ell_{p,q}$ regularization for (structured) sparsity. Finding solutions to these non-smooth and possibly non-convex problems typically relies on specialized optimization routines. In contrast, the method studied here is compatible with off-the-shelf (stochastic) gradient descent that is ubiquitous in deep learning, thereby enabling differentiable sparse regularization without approximations. The proposed optimization transfer comprises an overparametrization of selected model parameters followed by a change of penalties. In the overparametrized problem, smooth and convex $\ell_2$ regularization induces non-smooth and non-convex regularization in the original parametrization. We show that the resulting surrogate problem not only has an identical global optimum but also exactly preserves the local minima. This is particularly useful in non-convex regularization, where finding global solutions is NP-hard and local minima often generalize well. We provide an integrative overview that consolidates various literature strands on sparsity-inducing parametrizations in a general setting and meaningfully extend existing approaches. The feasibility of our approach is evaluated through numerical experiments, demonstrating its effectiveness by matching or outperforming common implementations of convex and non-convex regularizers.

Stochastic Gradient Descent (SGD) algorithms are widely used in optimizing neural networks, with Random Reshuffling (RR) and Single Shuffle (SS) being popular choices for cycling through random or single permutations of the training data. However, the convergence properties of these algorithms in the non-convex case are not fully understood. Existing results suggest that, in realistic training scenarios where the number of epochs is smaller than the training set size, RR may perform worse than SGD. In this paper, we analyze a general SGD algorithm that allows for arbitrary data orderings and show improved convergence rates for non-convex functions. Specifically, our analysis reveals that SGD with random and single shuffling is always faster or at least as good as classical SGD with replacement, regardless of the number of iterations. Overall, our study highlights the benefits of using SGD with random/single shuffling and provides new insights into its convergence properties for non-convex optimization.

In radial basis function neural network (RBFNN) based real-time learning tasks, forgetting mechanisms are widely used such that the neural network can keep its sensitivity to new data. However, with forgetting mechanisms, some useful knowledge will get lost simply because they are learned a long time ago, which we refer to as the passive knowledge forgetting phenomenon. To address this problem, this paper proposes a real-time training method named selective memory recursive least squares (SMRLS) in which the classical forgetting mechanisms are recast into a memory mechanism. Different from the forgetting mechanism, which mainly evaluates the importance of samples according to the time when samples are collected, the memory mechanism evaluates the importance of samples through both temporal and spatial distribution of samples. With SMRLS, the input space of the RBFNN is evenly divided into a finite number of partitions and a synthesized objective function is developed using synthesized samples from each partition. In addition to the current approximation error, the neural network also updates its weights according to the recorded data from the partition being visited. Compared with classical training methods including the forgetting factor recursive least squares (FFRLS) and stochastic gradient descent (SGD) methods, SMRLS achieves improved learning speed and generalization capability, which are demonstrated by corresponding simulation results.