This paper introduces PROMISE ($\textbf{Pr}$econditioned Stochastic $\textbf{O}$ptimization $\textbf{M}$ethods by $\textbf{I}$ncorporating $\textbf{S}$calable Curvature $\textbf{E}$stimates), a suite of sketching-based preconditioned stochastic gradient algorithms for solving large-scale convex optimization problems arising in machine learning. PROMISE includes preconditioned versions of SVRG, SAGA, and Katyusha; each algorithm comes with a strong theoretical analysis and effective default hyperparameter values. In contrast, traditional stochastic gradient methods require careful hyperparameter tuning to succeed, and degrade in the presence of ill-conditioning, a ubiquitous phenomenon in machine learning. Empirically, we verify the superiority of the proposed algorithms by showing that, using default hyperparameter values, they outperform or match popular tuned stochastic gradient optimizers on a test bed of $51$ ridge and logistic regression problems assembled from benchmark machine learning repositories. On the theoretical side, this paper introduces the notion of quadratic regularity in order to establish linear convergence of all proposed methods even when the preconditioner is updated infrequently. The speed of linear convergence is determined by the quadratic regularity ratio, which often provides a tighter bound on the convergence rate compared to the condition number, both in theory and in practice, and explains the fast global linear convergence of the proposed methods.
We consider trawl processes, which are stationary and infinitely divisible stochastic processes and can describe a wide range of statistical properties, such as heavy tails and long memory. In this paper, we develop the first likelihood-based methodology for the inference of real-valued trawl processes and introduce novel deterministic and probabilistic forecasting methods. Being non-Markovian, with a highly intractable likelihood function, trawl processes require the use of composite likelihood functions to parsimoniously capture their statistical properties. We formulate the composite likelihood estimation as a stochastic optimization problem for which it is feasible to implement iterative gradient descent methods. We derive novel gradient estimators with variances that are reduced by several orders of magnitude. We analyze both the theoretical properties and practical implementation details of these estimators and release a Python library which can be used to fit a large class of trawl processes. In a simulation study, we demonstrate that our estimators outperform the generalized method of moments estimators in terms of both parameter estimation error and out-of-sample forecasting error. Finally, we formalize a stochastic chain rule for our gradient estimators. We apply the new theory to trawl processes and provide a unified likelihood-based methodology for the inference of both real-valued and integer-valued trawl processes.
Quality-Diversity is a branch of stochastic optimization that is often applied to problems from the Reinforcement Learning and control domains in order to construct repertoires of well-performing policies/skills that exhibit diversity with respect to a behavior space. Such archives are usually composed of a finite number of reactive agents which are each associated to a unique behavior descriptor, and instantiating behavior descriptors outside of that coarsely discretized space is not straight-forward. While a few recent works suggest solutions to that issue, the trajectory that is generated is not easily customizable beyond the specification of a target behavior descriptor. We propose to jointly solve those problems in environments where semantic information about static scene elements is available by leveraging a Large Language Model to augment the repertoire with natural language descriptions of trajectories, and training a policy conditioned on those descriptions. Thus, our method allows a user to not only specify an arbitrary target behavior descriptor, but also provide the model with a high-level textual prompt to shape the generated trajectory. We also propose an LLM-based approach to evaluating the performance of such generative agents. Furthermore, we develop a benchmark based on simulated robot navigation in a 2d maze that we use for experimental validation.
Multi-party collaborative training, such as distributed learning and federated learning, is used to address the big data challenges. However, traditional multi-party collaborative training algorithms were mainly designed for balanced data mining tasks and are intended to optimize accuracy (\emph{e.g.}, cross-entropy). The data distribution in many real-world applications is skewed and classifiers, which are trained to improve accuracy, perform poorly when applied to imbalanced data tasks since models could be significantly biased toward the primary class. Therefore, the Area Under Precision-Recall Curve (AUPRC) was introduced as an effective metric. Although single-machine AUPRC maximization methods have been designed, multi-party collaborative algorithm has never been studied. The change from the single-machine to the multi-party setting poses critical challenges. To address the above challenge, we study the serverless multi-party collaborative AUPRC maximization problem since serverless multi-party collaborative training can cut down the communications cost by avoiding the server node bottleneck, and reformulate it as a conditional stochastic optimization problem in a serverless multi-party collaborative learning setting and propose a new ServerLess biAsed sTochastic gradiEnt (SLATE) algorithm to directly optimize the AUPRC. After that, we use the variance reduction technique and propose ServerLess biAsed sTochastic gradiEnt with Momentum-based variance reduction (SLATE-M) algorithm to improve the convergence rate, which matches the best theoretical convergence result reached by the single-machine online method. To the best of our knowledge, this is the first work to solve the multi-party collaborative AUPRC maximization problem.
In this study, we tackle the challenge of outlier-robust predictive modeling using highly expressive neural networks. Our approach integrates two key components: (1) a transformed trimmed loss (TTL), a computationally efficient variant of the classical trimmed loss, and (2) higher-order variation regularization (HOVR), which imposes smoothness constraints on the prediction function. While traditional robust statistics typically assume low-complexity models such as linear and kernel models, applying TTL alone to modern neural networks may fail to ensure robustness, as their high expressive power allows them to fit both inliers and outliers, even when a robust loss is used. To address this, we revisit the traditional notion of breakdown point and adapt it to the nonlinear function setting, introducing a regularization scheme via HOVR that controls the model's capacity and suppresses overfitting to outliers. We theoretically establish that our training procedure retains a high functional breakdown point, thereby ensuring robustness to outlier contamination. We develop a stochastic optimization algorithm tailored to this framework and provide a theoretical guarantee of its convergence.
Density power divergence (DPD) is designed to robustly estimate the underlying distribution of observations, in the presence of outliers. However, DPD involves an integral of the power of the parametric density models to be estimated; the explicit form of the integral term can be derived only for specific densities, such as normal and exponential densities. While we may perform a numerical integration for each iteration of the optimization algorithms, the computational complexity has hindered the practical application of DPD-based estimation to more general parametric densities. To address the issue, this study introduces a stochastic approach to minimize DPD for general parametric density models. The proposed approach can also be employed to minimize other density power-based $\gamma$-divergences, by leveraging unnormalized models. We provide \verb|R| package for implementation of the proposed approach in \url{https://github.com/oknakfm/sgdpd}.
Binary optimization has a wide range of applications in combinatorial optimization problems such as MaxCut, MIMO detection, and MaxSAT. However, these problems are typically NP-hard due to the binary constraints. We develop a novel probabilistic model to sample the binary solution according to a parameterized policy distribution. Specifically, minimizing the KL divergence between the parameterized policy distribution and the Gibbs distributions of the function value leads to a stochastic optimization problem whose policy gradient can be derived explicitly similar to reinforcement learning. For coherent exploration in discrete spaces, parallel Markov Chain Monte Carlo (MCMC) methods are employed to sample from the policy distribution with diversity and approximate the gradient efficiently. We further develop a filter scheme to replace the original objective function by the one with the local search technique to broaden the horizon of the function landscape. Convergence to stationary points in expectation of the policy gradient method is established based on the concentration inequality for MCMC. Numerical results show that this framework is very promising to provide near-optimal solutions for quite a few binary optimization problems.
With the increasing popularity and the increasing size of vision transformers (ViTs), there has been an increasing interest in making them more efficient and less computationally costly for deployment on edge devices with limited computing resources. Binarization can be used to help reduce the size of ViT models and their computational cost significantly, using popcount operations when the weights and the activations are in binary. However, ViTs suffer a larger performance drop when directly applying convolutional neural network (CNN) binarization methods or existing binarization methods to binarize ViTs compared to CNNs on datasets with a large number of classes such as ImageNet-1k. With extensive analysis, we find that binary vanilla ViTs such as DeiT miss out on a lot of key architectural properties that CNNs have that allow binary CNNs to have much higher representational capability than binary vanilla ViT. Therefore, we propose BinaryViT, in which inspired by the CNN architecture, we include operations from the CNN architecture into a pure ViT architecture to enrich the representational capability of a binary ViT without introducing convolutions. These include an average pooling layer instead of a token pooling layer, a block that contains multiple average pooling branches, an affine transformation right before the addition of each main residual connection, and a pyramid structure. Experimental results on the ImageNet-1k dataset show the effectiveness of these operations that allow a binary pure ViT model to be competitive with previous state-of-the-art (SOTA) binary CNN models.
We argue that inventory management presents unique opportunities for reliably applying and evaluating deep reinforcement learning (DRL). Toward reliable application, we emphasize and test two techniques. The first is Hindsight Differentiable Policy Optimization (HDPO), which performs stochastic gradient descent to optimize policy performance while avoiding the need to repeatedly deploy randomized policies in the environment-as is common with generic policy gradient methods. Our second technique involves aligning policy (neural) network architectures with the structure of the inventory network. Specifically, we focus on a network with a single warehouse that consolidates inventory from external suppliers, holds it, and then distributes it to many stores as needed. In this setting, we introduce the symmetry-aware policy network architecture. We motivate this architecture by establishing an asymptotic performance guarantee and empirically demonstrate its ability to reduce the amount of data needed to uncover strong policies. Both techniques exploit structures inherent in inventory management problems, moving beyond generic DRL algorithms. Toward rigorous evaluation, we create and share new benchmark problems, divided into two categories. One type focuses on problems with hidden structures that allow us to compute or bound the cost of the true optimal policy. Across four problems of this type, we find HDPO consistently attains near-optimal performance, handling up to 60-dimensional raw state vectors effectively. The other type of evaluation involves constructing a test problem using real time series data from a large retailer, where the optimum is poorly understood. Here, we find HDPO methods meaningfully outperform a variety of generalized newsvendor heuristics. Our code can be found at github.com/MatiasAlvo/Neural_inventory_control.
The training of over-parameterized neural networks has received much study in recent literature. An important consideration is the regularization of over-parameterized networks due to their highly nonconvex and nonlinear geometry. In this paper, we study noise injection algorithms, which can regularize the Hessian of the loss, leading to regions with flat loss surfaces. Specifically, by injecting isotropic Gaussian noise into the weight matrices of a neural network, we can obtain an approximately unbiased estimate of the trace of the Hessian. However, naively implementing the noise injection via adding noise to the weight matrices before backpropagation presents limited empirical improvements. To address this limitation, we design a two-point estimate of the Hessian penalty, which injects noise into the weight matrices along both positive and negative directions of the random noise. In particular, this two-point estimate eliminates the variance of the first-order Taylor's expansion term on the Hessian. We show a PAC-Bayes generalization bound that depends on the trace of the Hessian (and the radius of the weight space), which can be measured from data. We conduct a detailed experimental study to validate our approach and show that it can effectively regularize the Hessian and improve generalization. First, our algorithm can outperform prior approaches on sharpness-reduced training, delivering up to a 2.4% test accuracy increase for fine-tuning ResNets on six image classification datasets. Moreover, the trace of the Hessian reduces by 15.8%, and the largest eigenvalue is reduced by 9.7% with our approach. We also find that the regularization of the Hessian can be combined with weight decay and data augmentation, leading to stronger regularization. Second, our approach remains effective for improving generalization in pretraining multimodal CLIP models and chain-of-thought fine-tuning.
Identifying clients with similar objectives and learning a model-per-cluster is an intuitive and interpretable approach to personalization in federated learning. However, doing so with provable and optimal guarantees has remained an open challenge. We formalize this problem as a stochastic optimization problem, achieving optimal convergence rates for a large class of loss functions. We propose simple iterative algorithms which identify clusters of similar clients and train a personalized model-per-cluster, using local client gradients and flexible constraints on the clusters. The convergence rates of our algorithms asymptotically match those obtained if we knew the true underlying clustering of the clients and are provably robust in the Byzantine setting where some fraction of the clients are malicious.
We develop a principled approach to end-to-end learning in stochastic optimization. First, we show that the standard end-to-end learning algorithm admits a Bayesian interpretation and trains a posterior Bayes action map. Building on the insights of this analysis, we then propose new end-to-end learning algorithms for training decision maps that output solutions of empirical risk minimization and distributionally robust optimization problems, two dominant modeling paradigms in optimization under uncertainty. Numerical results for a synthetic newsvendor problem illustrate the key differences between alternative training schemes. We also investigate an economic dispatch problem based on real data to showcase the impact of the neural network architecture of the decision maps on their test performance.
Second order stochastic optimizers allow parameter update step size and direction to adapt to loss curvature, but have traditionally required too much memory and compute for deep learning. Recently, Shampoo [Gupta et al., 2018] introduced a Kronecker factored preconditioner to reduce these requirements: it is used for large deep models [Anil et al., 2020] and in production [Anil et al., 2022]. However, it takes inverse matrix roots of ill-conditioned matrices. This requires 64-bit precision, imposing strong hardware constraints. In this paper, we propose a novel factorization, Kronecker Approximation-Domination (KrAD). Using KrAD, we update a matrix that directly approximates the inverse empirical Fisher matrix (like full matrix AdaGrad), avoiding inversion and hence 64-bit precision. We then propose KrADagrad$^\star$, with similar computational costs to Shampoo and the same regret. Synthetic ill-conditioned experiments show improved performance over Shampoo for 32-bit precision, while for several real datasets we have comparable or better generalization.
We introduce a novel approach for analyzing the training dynamics of ReLU networks by examining the characteristic activation boundaries of individual ReLU neurons. Our proposed analysis reveals a critical instability in common neural network parameterizations and normalizations during stochastic optimization, which impedes fast convergence and hurts generalization performance. Addressing this, we propose Geometric Parameterization (GmP), a novel neural network parameterization technique that effectively separates the radial and angular components of weights in the hyperspherical coordinate system. We show theoretically that GmP resolves the aforementioned instability issue. We report empirical results on various models and benchmarks to verify GmP's advantages of optimization stability, convergence speed and generalization performance.
With the rapid growth in model size, fine-tuning the large pre-trained language model has become increasingly difficult due to its extensive memory usage. Previous works usually focus on reducing the number of trainable parameters in the network. While the model parameters do contribute to memory usage, the primary memory bottleneck during training arises from storing feature maps, also known as activations, as they are crucial for gradient calculation. Notably, neural networks are usually trained using stochastic gradient descent. We argue that in stochastic optimization, models can handle noisy gradients as long as the gradient estimator is unbiased with reasonable variance. Following this motivation, we propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance, which only requires storing the sub-sampled activations for calculating the gradient. Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones. By replacing the linear operation with our approximated one in transformers, we can achieve up to 2.7$\times$ peak memory reduction with almost no accuracy drop and enables up to $6.4\times$ larger batch size. Under the same hardware, WTA-CRS enables better down-streaming task performance by applying larger models and/or faster training speed with larger batch sizes.
This paper considers a novel application of deep AUC maximization (DAM) for multi-instance learning (MIL), in which a single class label is assigned to a bag of instances (e.g., multiple 2D slices of a CT scan for a patient). We address a neglected yet non-negligible computational challenge of MIL in the context of DAM, i.e., bag size is too large to be loaded into {GPU} memory for backpropagation, which is required by the standard pooling methods of MIL. To tackle this challenge, we propose variance-reduced stochastic pooling methods in the spirit of stochastic optimization by formulating the loss function over the pooled prediction as a multi-level compositional function. By synthesizing techniques from stochastic compositional optimization and non-convex min-max optimization, we propose a unified and provable muli-instance DAM (MIDAM) algorithm with stochastic smoothed-max pooling or stochastic attention-based pooling, which only samples a few instances for each bag to compute a stochastic gradient estimator and to update the model parameter. We establish a similar convergence rate of the proposed MIDAM algorithm as the state-of-the-art DAM algorithms. Our extensive experiments on conventional MIL datasets and medical datasets demonstrate the superiority of our MIDAM algorithm.
Training a modern machine learning architecture on a new task requires extensive learning-rate tuning, which comes at a high computational cost. Here we develop new Polyak-type adaptive learning rates that can be used on top of any momentum method, and require less tuning to perform well. We first develop MoMo, a Momentum Model based adaptive learning rate for SGD-M (stochastic gradient descent with momentum). MoMo uses momentum estimates of the losses and gradients sampled at each iteration to build a model of the loss function. Our model makes use of any known lower bound of the loss function by using truncation, e.g. most losses are lower-bounded by zero. The model is then approximately minimized at each iteration to compute the next step. We show how MoMo can be used in combination with any momentum-based method, and showcase this by developing MoMo-Adam, which is Adam with our new model-based adaptive learning rate. We show that MoMo attains a $\mathcal{O}(1/\sqrt{K})$ convergence rate for convex problems with interpolation, needing knowledge of no problem-specific quantities other than the optimal value. Additionally, for losses with unknown lower bounds, we develop on-the-fly estimates of a lower bound, that are incorporated in our model. We show that MoMo and MoMo-Adam improve over SGD-M and Adam in terms of robustness to hyperparameter tuning for training image classifiers on MNIST, CIFAR, and Imagenet, for recommender systems on Criteo, for a transformer model on the translation task IWSLT14, and for a diffusion model.
Diffusion models have emerged as a key pillar of foundation models in visual domains. One of their critical applications is to universally solve different downstream inverse tasks via a single diffusion prior without re-training for each task. Most inverse tasks can be formulated as inferring a posterior distribution over data (e.g., a full image) given a measurement (e.g., a masked image). This is however challenging in diffusion models since the nonlinear and iterative nature of the diffusion process renders the posterior intractable. To cope with this challenge, we propose a variational approach that by design seeks to approximate the true posterior distribution. We show that our approach naturally leads to regularization by denoising diffusion process (RED-Diff) where denoisers at different timesteps concurrently impose different structural constraints over the image. To gauge the contribution of denoisers from different timesteps, we propose a weighting mechanism based on signal-to-noise-ratio (SNR). Our approach provides a new variational perspective for solving inverse problems with diffusion models, allowing us to formulate sampling as stochastic optimization, where one can simply apply off-the-shelf solvers with lightweight iterates. Our experiments for image restoration tasks such as inpainting and superresolution demonstrate the strengths of our method compared with state-of-the-art sampling-based diffusion models.
Graph neural networks (GNN) suffer from severe inefficiency. It is mainly caused by the exponential growth of node dependency with the increase of layers. It extremely limits the application of stochastic optimization algorithms so that the training of GNN is usually time-consuming. To address this problem, we propose to decouple a multi-layer GNN as multiple simple modules for more efficient training, which is comprised of classical forward training (FT)and designed backward training (BT). Under the proposed framework, each module can be trained efficiently in FT by stochastic algorithms without distortion of graph information owing to its simplicity. To avoid the only unidirectional information delivery of FT and sufficiently train shallow modules with the deeper ones, we develop a backward training mechanism that makes the former modules perceive the latter modules. The backward training introduces the reversed information delivery into the decoupled modules as well as the forward information delivery. To investigate how the decoupling and greedy training affect the representational capacity, we theoretically prove that the error produced by linear modules will not accumulate on unsupervised tasks in most cases. The theoretical and experimental results show that the proposed framework is highly efficient with reasonable performance.
In data-driven stochastic optimization, model parameters of the underlying distribution need to be estimated from data in addition to the optimization task. Recent literature considers integrating the estimation and optimization processes by selecting model parameters that lead to the best empirical objective performance. This integrated approach, which we call integrated-estimation-optimization (IEO), can be readily shown to outperform simple estimate-then-optimize (ETO) when the model is misspecified. In this paper, we show that a reverse behavior appears when the model class is well-specified and there is sufficient data. Specifically, for a general class of nonlinear stochastic optimization problems, we show that simple ETO outperforms IEO asymptotically when the model class covers the ground truth, in the strong sense of stochastic dominance of the regret. Namely, the entire distribution of the regret, not only its mean or other moments, is always better for ETO compared to IEO. Our results also apply to constrained, contextual optimization problems where the decision depends on observed features. Whenever applicable, we also demonstrate how standard sample average approximation (SAA) performs the worst when the model class is well-specified in terms of regret, and best when it is misspecified. Finally, we provide experimental results to support our theoretical comparisons and illustrate when our insights hold in finite-sample regimes and under various degrees of misspecification.