Sparsity Acceleration

Previous studies yielded discouraging results for item-level locally differentially private linear regression with $s$-sparsity assumption, where the minimax rate for $nm$ samples is $\mathcal{O}(sd / nm\varepsilon^2)$. This can be challenging for high-dimensional data, where the dimension $d$ is extremely large. In this work, we investigate user-level locally differentially private sparse linear regression. We show that with $n$ users each contributing $m$ samples, the linear dependency of dimension $d$ can be eliminated, yielding an error upper bound of $\mathcal{O}(s/ nm\varepsilon^2)$. We propose a framework that first selects candidate variables and then conducts estimation in the narrowed low-dimensional space, which is extendable to general sparse estimation problems with tight error bounds. Experiments on both synthetic and real datasets demonstrate the superiority of the proposed methods. Both the theoretical and empirical results suggest that, with the same number of samples, locally private sparse estimation is better conducted when multiple samples per user are available.

The sparse Johnson-Lindenstrauss transform is one of the central techniques in dimensionality reduction. It supports embedding a set of $n$ points in $\mathbb{R}^d$ into $m=O(\varepsilon^{-2} \ln n)$ dimensions while preserving all pairwise distances to within $1 \pm \varepsilon$. Each input point $x$ is embedded to $Ax$, where $A$ is an $m \times d$ matrix having $s$ non-zeros per column, allowing for an embedding time of $O(s \|x\|_0)$. Since the sparsity of $A$ governs the embedding time, much work has gone into improving the sparsity $s$. The current state-of-the-art by Kane and Nelson (2014) shows that $s = O(\varepsilon^{-1} \ln n)$ suffices. This is almost matched by a lower bound of $s = \Omega(\varepsilon^{-1} \ln n/\ln(1/\varepsilon))$ by Nelson and Nguyen (2013) for $d=\Omega(n)$. Previous work thus suggests that we have near-optimal embeddings. In this work, we revisit sparse embeddings and present a sparser embedding for instances in which $d = n^{o(1)}$, which in many applications is realistic. Formally, our embedding achieves $s = O(\varepsilon^{-1}(\ln n/\ln(1/\varepsilon)+\ln^{2/3}n \ln^{1/3} d))$. We also complement our analysis by strengthening the lower bound of Nelson and Nguyen to hold also when $d \ll n$, thereby matching the first term in our new sparsity upper bound. Finally, we also improve the sparsity of the best oblivious subspace embeddings for optimal embedding dimensionality.

Static sparse training aims to train sparse models from scratch, achieving remarkable results in recent years. A key design choice is given by the sparse initialization, which determines the trainable sub-network through a binary mask. Existing methods mainly select such mask based on a predefined dense initialization. Such an approach may not efficiently leverage the mask's potential impact on the optimization. An alternative direction, inspired by research into dynamical isometry, is to introduce orthogonality in the sparse subnetwork, which helps in stabilizing the gradient signal. In this work, we propose Exact Orthogonal Initialization (EOI), a novel sparse orthogonal initialization scheme based on composing random Givens rotations. Contrary to other existing approaches, our method provides exact (not approximated) orthogonality and enables the creation of layers with arbitrary densities. We demonstrate the superior effectiveness and efficiency of EOI through experiments, consistently outperforming common sparse initialization techniques. Our method enables training highly sparse 1000-layer MLP and CNN networks without residual connections or normalization techniques, emphasizing the crucial role of weight initialization in static sparse training alongside sparse mask selection.

The transformer architecture has prevailed in various deep learning settings due to its exceptional capabilities to select and compose structural information. Motivated by these capabilities, Sanford et al. (2023) proposed the *sparse token selection* task, in which transformers excel while fully-connected networks (FCNs) fail in the worst case. Building upon that, we strengthen the FCN lower bound to an average-case setting and establish an algorithmic separation of transformers over FCNs. Specifically, a one-layer transformer trained with gradient descent provably learns the sparse token selection task and, surprisingly, exhibits strong out-of-distribution length generalization. We provide empirical simulations to justify our theoretical findings.

It is known that sparsity can improve interpretability for deep neural networks. However, existing methods in the area either require networks that are pre-trained with sparsity constraints, or impose sparsity after the fact, altering the network's general behavior. In this paper, we demonstrate, for the first time, that sparsity can instead be incorporated into the interpretation process itself, as a sample-specific preprocessing step. Unlike previous work, this approach, which we call SPADE, does not place constraints on the trained model and does not affect its behavior during inference on the sample. Given a trained model and a target sample, SPADE uses sample-targeted pruning to provide a "trace" of the network's execution on the sample, reducing the network to the most important connections prior to computing an interpretation. We demonstrate that preprocessing with SPADE significantly increases the accuracy of image saliency maps across several interpretability methods. Additionally, SPADE improves the usefulness of neuron visualizations, aiding humans in reasoning about network behavior. Our code is available at https://github.com/IST-DASLab/SPADE.

Training large transformers is slow, but recent innovations on GPU architecture give us an advantage. NVIDIA Ampere GPUs can execute a fine-grained 2:4 sparse matrix multiplication twice as fast as its dense equivalent. In the light of this property, we comprehensively investigate the feasibility of accelerating feed-forward networks (FFNs) of transformers in pre-training. First, we define a ``flip rate'' to monitor the stability of a 2:4 training process. Utilizing this metric, we propose three techniques to preserve accuracy: to modify the sparse-refined straight-through estimator by applying the masked decay term on gradients, to determine a feasible decay factor in warm-up stage, and to enhance the model's quality by a dense fine-tuning procedure near the end of pre-training. Besides, we devise two techniques to practically accelerate training: to calculate transposable 2:4 masks by convolution, and to accelerate gated activation functions by reducing GPU L2 cache miss. Experiments show that our 2:4 sparse training algorithm achieves similar convergence to dense training algorithms on several transformer pre-training tasks, while actual acceleration can be observed on different shapes of transformer block apparently. Our toolkit is available at https://github.com/huyz2023/2by4-pretrain.

We propose a novel algorithmic framework for distributional reinforcement learning, based on learning finite-dimensional mean embeddings of return distributions. The framework reveals a wide variety of new algorithms for dynamic programming and temporal-difference algorithms that rely on the sketch Bellman operator, which updates mean embeddings with simple linear-algebraic computations. We provide asymptotic convergence theory, and examine the empirical performance of the algorithms on a suite of tabular tasks. Further, we show that this approach can be straightforwardly combined with deep reinforcement learning.

Block coordinate descent is a powerful algorithmic template suitable for big data optimization. This template admits a lot of variants including block gradient descent (BGD), which performs gradient descent on a selected block of variables, while keeping other variables fixed. For a very long time, the stepsize for each block has tacitly been set to one divided by the block-wise Lipschitz smoothness constant, imitating the vanilla stepsize rule for gradient descent (GD). However, such a choice for BGD has not yet been able to theoretically justify its empirical superiority over GD, as existing convergence rates for BGD have worse constants than GD in the deterministic cases. To discover such theoretical justification, we set up a simple environment where we consider BGD applied to least-squares with two blocks of variables. Assuming the data matrix corresponding to each block is orthogonal, we find optimal stepsizes of BGD in closed form, which provably lead to asymptotic convergence rates twice as fast as GD with Polyak's momentum; this means, under that orthogonality assumption, one can accelerate BGD by just tuning stepsizes and without adding any momentum. An application that satisfies this assumption is *generalized alternating projection* between two subspaces, and applying our stepsizes to it improves the prior convergence rate that was once claimed, slightly inaccurately, to be optimal. The main proof idea is to minimize, in stepsize variables, the spectral radius of a matrix that controls convergence rates.

Unbalanced optimal transport (UOT) has recently gained much attention due to its flexible framework for handling un-normalized measures and its robustness properties. In this work, we explore learning (structured) sparse transport plans in the UOT setting, i.e., transport plans have an upper bound on the number of non-sparse entries in each column (structured sparse pattern) or in the whole plan (general sparse pattern). We propose novel sparsity-constrained UOT formulations building on the recently explored maximum mean discrepancy based UOT. We show that the proposed optimization problem is equivalent to the maximization of a weakly submodular function over a uniform matroid or a partition matroid. We develop efficient gradient-based discrete greedy algorithms and provide the corresponding theoretical guarantees. Empirically, we observe that our proposed greedy algorithms select a diverse support set and we illustrate the efficacy of the proposed approach in various applications.

Large Language Models (LLMs), renowned for their remarkable performance across diverse domains, present a challenge due to their colossal model size when it comes to practical deployment. In response to this challenge, efforts have been directed toward the application of traditional network pruning techniques to LLMs, uncovering a massive number of parameters can be pruned in one-shot without hurting performance. Building upon insights gained from pre-LLM models, particularly BERT-level language models, prevailing LLM pruning strategies have consistently adhered to the practice of uniformly pruning all layers at equivalent sparsity levels, resulting in robust performance. However, this observation stands in contrast to the prevailing trends observed in the field of vision models, where non-uniform layerwise sparsity typically yields substantially improved results. To elucidate the underlying reasons for this disparity, we conduct a comprehensive analysis of the distribution of token features within LLMs. In doing so, we discover a strong correlation with the emergence of outliers, defined as features exhibiting significantly greater magnitudes compared to their counterparts in feature dimensions. Inspired by this finding, we introduce a novel LLM pruning methodology that incorporates a tailored set of **non-uniform layerwise sparsity ratios** specifically designed for LLM pruning, termed as **O**utlier **W**eighed **L**ayerwise sparsity (**OWL**). The sparsity ratio of OWL is directly proportional to the outlier ratio observed within each layer, facilitating a more effective alignment between layerwise weight sparsity and outlier ratios. Our empirical evaluation, conducted across the LLaMA-V1/V2, Vicuna, OPT, and Mistral, spanning various benchmarks, demonstrates the distinct advantages offered by OWL over previous methods. For instance, OWL exhibits a remarkable performance gain, surpassing the state-of-the-art Wanda and SparseGPT by **61.22** and **6.80** perplexity at a high sparsity level of 70%, respectively, while delivering **2.6$\times$** end-to-end inference speed-up in the DeepSparse inference engine. Code is available at https://github.com/luuyin/OWL.git.

Recent research has focused on weight sparsity in deep neural network training to reduce FLOPs, aiming for improved efficiency (test accuracy w.r.t training FLOPs). However, sparse weight training often compromises accuracy, requiring extended training schedules to attain the accuracy of dense models. In contrast, our approach, Sparse Iso-FLOP Transformations (Sparse-IFT), uses sparsity to improve accuracy while maintaining dense model FLOPs. Using a single hyperparameter (i.e., the sparsity level), Sparse-IFTs efficiently replace dense layers, expanding the search space for optimal sparse masks. In addition, dynamic sparse training (DST) with Sparse-IFT models effectively navigate this larger sparse mask-weight space, which is evidenced by a spectral analysis using Ramanujan graph properties. Our study reveals a robust correlation among mask topology, weights, and final performance. Notably, without adjusting any training hyperparameters, replacing dense layers with Sparse-IFT yields significant improvements, such as a +3.5% boost for ResNet-18 on ImageNet and +0.9% for GPT-3 Small on the Open LLM leaderboard. To the best of our knowledge, this is the first work to demonstrate the use of sparsity for improving the accuracy of dense models through a set of simple-to-use sparse transformations. Code is available at: https://github.com/CerebrasResearch/Sparse-IFT.

Spiking Neural Networks (SNNs) have attracted great attention for their energy-efficient operations and biologically inspired structures, offering potential advantages over Artificial Neural Networks (ANNs) in terms of energy efficiency and interpretability. Nonetheless, similar to ANNs, the robustness of SNNs remains a challenge, especially when facing adversarial attacks. Existing techniques, whether adapted from ANNs or specifically designed for SNNs, exhibit limitations in training SNNs or defending against strong attacks. In this paper, we propose a novel approach to enhance the robustness of SNNs through gradient sparsity regularization. We observe that SNNs exhibit greater resilience to random perturbations compared to adversarial perturbations, even at larger scales. Motivated by this, we aim to narrow the gap between SNNs under adversarial and random perturbations, thereby improving their overall robustness. To achieve this, we theoretically prove that this performance gap is upper bounded by the gradient sparsity of the probability associated with the true label concerning the input image, laying the groundwork for a practical strategy to train robust SNNs by regularizing the gradient sparsity. We validate the effectiveness of our approach through extensive experiments on both image-based and event-based datasets. The results demonstrate notable improvements in the robustness of SNNs. Our work highlights the importance of gradient sparsity in SNNs and its role in enhancing robustness.

Sparse Neural Networks (SNNs) have received voluminous attention for mitigating the explosion in computational costs and memory footprints of modern deep neural networks. Despite their popularity, most state-of-the-art training approaches seek to find a single high-quality sparse subnetwork with a preset sparsity pattern and ratio, making them inadequate to satiate platform and resource variability. Recently proposed approaches attempt to jointly train multiple subnetworks (we term as ``sparse co-training") with a *fixed sparsity pattern*, to allow switching sparsity ratios subject to resource requirements. In this work, we take one more step forward and expand the scope of sparse co-training to cover diverse sparsity patterns and multiple sparsity ratios *at once*. We introduce **Sparse Cocktail**, the first sparse co-training framework that co-trains a suite of sparsity patterns simultaneously, loaded with multiple sparsity ratios which facilitate harmonious switch across various sparsity patterns and ratios at inference depending on the hardware availability. More specifically, Sparse Cocktail alternatively trains subnetworks generated from different sparsity patterns with a gradual increase in sparsity ratios across patterns and relies on an *unified mask generation process* and the *Dense Pivot Co-training* to ensure the subnetworks of different patterns orchestrate their shared parameters without canceling each other’s performance. Experiment results on image classification, object detection, and instance segmentation illustrate the favorable effectiveness and flexibility of Sparse Cocktail, pointing to a promising direction for sparse co-training. Codes will be released.

The *straight-through estimator* (STE) is commonly used to optimize quantized neural networks, yet its contexts of effective performance are still unclear despite empirical successes. To make a step forward in this comprehension, we apply STE to a well-understood problem: *sparse support recovery*. We introduce the *Support Exploration Algorithm* (SEA), a novel algorithm promoting sparsity, and we analyze its performance in support recovery (a.k.a. model selection) problems. SEA explores more supports than the state-of-the-art, leading to superior performance in experiments, especially when the columns of $A$ are strongly coherent. The theoretical analysis considers recovery guarantees when the linear measurements matrix $A$ satisfies the *Restricted Isometry Property* (RIP). The sufficient conditions of recovery are comparable but more stringent than those of the state-of-the-art in sparse support recovery. Their significance lies mainly in their applicability to an instance of the STE.

Pre-trained Transformers inherently possess the characteristic of sparse activation, where only a small fraction of the neurons are activated for each token. While sparse activation has been explored through post-training methods, its potential in pre-training remains untapped. In this work, we first study how activation properties change during pre-training. Our examination reveals that Transformers exhibit sparse activation throughout the majority of the pre-training process while the activation correlation keeps evolving as training progresses. Leveraging this observation, we propose Switchable Sparse-Dense Learning (SSD). SSD adaptively switches between the Mixtures-of-Experts (MoE) based sparse training and the conventional dense training during the pre-training process, leveraging the efficiency of sparse training and avoiding the static activation correlation of sparse training. Compared to dense training, SSD achieves comparable performance with identical model size and reduces pre-training costs. Moreover, the models trained with SSD can be directly used as MoE models for sparse inference and achieve the same performance as dense models with up to $2\times$ faster inference speed. Codes are available at https://github.com/thunlp/moefication.

The remarkable success of modern machine learning models on large datasets often demands extensive training time and resource consumption. To save cost, a prevalent research line, known as online batch selection, explores selecting informative subsets during the training process. Although recent efforts achieve advancements by measuring the impact of each sample on generalization, their reliance on additional reference models inherently limits their practical applications, when there are no such ideal models available. On the other hand, the vanilla reference-model-free methods involve independently scoring and selecting data in a sample-wise manner, which sacrifices the diversity and induces the redundancy. To tackle this dilemma, we propose Diversified Batch Selection (DivBS), which is reference-model-free and can efficiently select diverse and representative samples. Specifically, we define a novel selection objective that measures the group-wise orthogonalized representativeness to combat the redundancy issue of previous sample-wise criteria, and provide a principled selection-efficient realization. Extensive experiments across various tasks demonstrate the significant superiority of DivBS in the performance-speedup trade-off. The code is publicly available.

We study the contextual bandits with knapsack (CBwK) problem under the high-dimensional setting where the dimension of the feature is large. We investigate how to exploit the sparsity structure to achieve improved regret for the CBwK problem. To this end, we first develop an online variant of the hard thresholding algorithm that performs the optimal sparse estimation. We further combine our online estimator with a primal-dual framework, where we assign a dual variable to each knapsack constraint and utilize an online learning algorithm to update the dual variable, thereby controlling the consumption of the knapsack capacity. We show that this integrated approach allows us to achieve a sublinear regret that depends logarithmically on the feature dimension, thus improving the polynomial dependency established in the previous literature. We also apply our framework to the high-dimension contextual bandit problem without the knapsack constraint and achieve optimal regret in both the data-poor regime and the data-rich regime.

With the prevalence of pre-training-fine-tuning paradigm, how to efficiently adapt the pre-trained model to the downstream tasks has been an intriguing issue. $\textbf{P}$arameter-$\textbf{E}$fficient $\textbf{F}$ine-$\textbf{T}$uning(PEFT) methods have been proposed for low-cost adaptation. Although PEFT has demonstrated effectiveness and been widely applied, the underlying principles are still unclear. In this paper, we adopt the PAC-Bayesian generalization error bound, viewing pre-training as a shift of prior distribution which leads to a tighter bound for generalization error. We validate this shift from the perspectives of oscillations in the loss landscape and the quasi-sparsity in gradient distribution. Based on this, we propose a gradient-based sparse fine-tuning algorithm, named $\textbf{S}$parse $\textbf{I}$ncrement $\textbf{F}$ine-$\textbf{T}$uning(SIFT), and validate its effectiveness on a range of tasks including the GLUE Benchmark and Instruction-tuning. The code is accessible at https://github.com/song-wx/SIFT/.