ICLR 2025 论文列表

Estimating matrices in the symmetric positive-definite (SPD) cone is of interest for many applications ranging from computer vision to graph learning. While there exist various convex optimization-based estimators, they remain limited in expressivity due to their model-based approach. The success of deep learning motivates the use of learning-based approaches to estimate SPD matrices with neural networks in a data-driven fashion. However, designing effective neural architectures for SPD learning is challenging, particularly when the task requires additional structural constraints, such as element-wise sparsity. Current approaches either do not ensure that the output meets all desired properties or lack expressivity. In this paper, we introduce SpodNet, a novel and generic learning module that guarantees SPD outputs and supports additional structural constraints. Notably, it solves the challenging task of learning jointly SPD and sparse matrices. Our experiments illustrate the versatility and relevance of SpodNet layers for such applications.

Despite the efficacy of network sparsity in alleviating the deployment strain of Large Language Models (LLMs), it endures significant performance degradation. Applying Low-Rank Adaptation (LoRA) to fine-tune the sparse LLMs offers an intuitive approach to counter this predicament, while it holds shortcomings include: 1) The inability to integrate LoRA weights into sparse LLMs post-training, and 2) Insufficient performance recovery at high sparsity ratios. In this paper, we introduces dynamic $\textbf{Lo}$w-rank $\textbf{S}$parse $\textbf{A}$daptation $\textbf{(LoSA)}$, a novel method that seamlessly integrates low-rank adaptation into LLM sparsity within a unified framework, thereby enhancing the performance of sparse LLMs without increasing the inference latency. In particular, LoSA dynamically sparsifies the LoRA outcomes based on the corresponding sparse weights during fine-tuning, thus guaranteeing that the LoRA module can be integrated into the sparse LLMs post-training. Besides, to achieve the optimal sparse model architecture, LoSA leverages Representation Mutual Information (RMI) as an indicator to determine the importance of layers, thereby dynamically determining the optimal layer-wise sparsity rates during fine-tuning. Predicated on this, LoSA adjusts the rank of the LoRA module based on the variability in layer-wise reconstruction errors, allocating an appropriate fine-tuning for each layer to reduce the output discrepancies between dense and sparse LLMs. Extensive experiments tell that LoSA can efficiently boost the efficacy of sparse LLMs within a few hours, without introducing any additional inferential burden. For example, LoSA reduced the perplexity of sparse LLaMA-2-7B by $\textbf{68.73}$$\downarrow$ and increased zero-shot accuracy by $\textbf{16.32}$%$\uparrow$, achieving a $\textbf{2.60$\times$}$ speedup on CPU and $\textbf{2.23$\times$}$ speedup on GPU, requiring only $\textbf{45 minutes}$ of fine-tuning on $\textbf{a single}$ NVIDIA A100 80GB GPU. Code is available at https://github.com/wzhuang-xmu/LoSA.

The recent work by Dong and Yang (2023) showed for misspecified sparse linear bandits, one can obtain an $O(\epsilon)$-optimal policy using a polynomial number of samples when the sparsity is a constant, where $\epsilon$ is the misspecification error. This result is in sharp contrast to misspecified linear bandits without sparsity, which require an exponential number of samples to get the same guarantee. In order to study whether the analog result is possible in the reinforcement learning setting, we consider the following problem: assuming the optimal $Q$-function is a $d$-dimensional linear function with sparsity $k$ and misspecification error $\epsilon$, whether we can obtain an $O(\epsilon)$-optimal policy using number of samples polynomially in the feature dimension $d$. We first demonstrate why the standard approach based on Bellman backup or the existing optimistic value function elimination approach such as OLIVE (Jiang et al., 2017) achieves suboptimal guarantees for this problem. We then design a novel elimination-based algorithm to show one can obtain an $O(H\epsilon)$-optimal policy with sample complexity polynomially in the feature dimension $d$ and planning horizon $H$. Lastly, we complement our upper bound with an $\tilde \Omega(H\epsilon)$ suboptimality lower bound, giving a complete picture of this problem.

Zeroth-order optimization (ZO) is a memory-efficient strategy for fine-tuning Large Language Models using only forward passes. However, applying ZO fine-tuning in memory-constrained settings such as mobile phones and laptops remains challenging since these settings often involve weight quantization, while ZO requires full-precision perturbation and update. In this study, we address this limitation by combining static sparse ZO fine-tuning with quantization. Our approach transfers a small, static subset (0.1%) of "sensitive" parameters from pre-training to downstream tasks, focusing fine-tuning on this sparse set of parameters. The remaining untuned parameters are quantized, reducing memory demands. Our proposed workflow enables efficient ZO fine-tuning of an Llama2-7B model on a GPU device with less than 8GB of memory while outperforming full model ZO fine-tuning performance and in-context learning.

Pruning at Initialization (PaI) is a technique in neural network optimization characterized by the proactive elimination of weights before the network's training on designated tasks. This innovative strategy potentially reduces the costs for training and inference, significantly advancing computational efficiency. A key factor leading to PaI's effectiveness is that it considers the saliency of weights in an untrained network, and prioritizes the trainability and optimization potential of the pruned subnetworks. Recent methods can effectively prevent the formation of hard-to-optimize networks, e.g. through iterative adjustments at each network layer. However, this way often results in large-scale discrete optimization problems, which could make PaI further challenging. This paper introduces a novel method, called DPaI, that involves a differentiable optimization of the pruning mask. DPaI adopts a dynamic and adaptable pruning process, allowing easier optimization processes and better solutions. More importantly, our differentiable formulation enables readily use of the existing rich body of efficient gradient-based methods for PaI. Our empirical results demonstrate that DPaI significantly outperforms current state-of-the-art PaI methods on various architectures, such as Convolutional Neural Networks and Vision-Transformers. Code is available at https://github.com/QuanNguyen-Tri/DPaI.git

We study beyond worst-case dimensionality reduction for $s$-sparse vectors (vectors with at most $s$ non-zero coordinates). Our work is divided into two parts, each focusing on a different facet of beyond worst-case analysis: \noindent (a) We first consider average-case guarantees for embedding $s$-sparse vectors. Here, a well-known folklore upper bound based on the birthday-paradox states: For any collection $X$ of $s$-sparse vectors in $\mathbb{R}^d$, there exists a linear map $A: \mathbb{R}^d \rightarrow \mathbb{R}^{O(s^2)}$ which \emph{exactly} preserves the norm of $99\%$ of the vectors in $X$ in any $\ell_p$ norm (as opposed to the usual setting where guarantees hold for all vectors). We provide novel lower bounds showing that this is indeed optimal in many settings. Specifically, any oblivious linear map satisfying similar average-case guarantees must map to $\Omega(s^2)$ dimensions. The same lower bound also holds for a wider class of sufficiently smooth maps, including `encoder-decoder schemes', where we compare the norm of the original vector to that of a smooth function of the embedding. These lower bounds reveal a surprising separation result for smooth embeddings of sparse vectors, as an upper bound of $O(s \log(d))$ is possible if we instead use arbitrary functions, e.g., via compressed sensing algorithms. (b) Given these lower bounds, we specialize to sparse \emph{non-negative} vectors to hopes of improved upper bounds. For a dataset $X$ of non-negative $s$-sparse vectors and any $p \ge 1$, we can non-linearly embed $X$ to $O(s\log(|X|s)/\varepsilon^2)$ dimensions while preserving all pairwise distances in $\ell_p$ norm up to $1\pm \varepsilon$, with no dependence on $p$. Surprisingly, the non-negativity assumption enables much smaller embeddings than arbitrary sparse vectors, where the best known bound suffers an exponential $(\log |X|)^{O(p)}$ dependence. Our map also guarantees \emph{exact} dimensionality reduction for the $\ell_{\infty}$ norm by embedding $X$ into $O(s\log |X|)$ dimensions, which is tight. We further give separation results showing that both the non-linearity of $f$ and the non-negativity of $X$ are necessary, and provide downstream algorithmic improvements using our embedding.

Activation sparsity can enable practical inference speedups in large language models (LLMs) by reducing the compute and memory-movement required for matrix multiplications during the forward pass. However, existing methods face limitations that inhibit widespread adoption. Some approaches are tailored towards older models with ReLU-based sparsity, while others require extensive continued pre-training on up to hundreds of billions of tokens. This paper describes TEAL (**T**raining-Fre**e** **A**ctivation Sparsity in **L**LMs), a simple training-free method that applies magnitude-based activation sparsity to hidden states throughout the entire model. TEAL achieves 40-50\% model-wide sparsity with minimal performance degradation across Llama-2, Llama-3, and Mistral families, with sizes varying from 7B to 70B. We improve existing sparse kernels and demonstrate wall-clock decoding speed-ups of up to 1.53× and 1.8× at 40\% and 50\% model-wide sparsity. TEAL is compatible with weight quantization, enabling further efficiency gains.

The human brain excels at lifelong learning by not only encoding information in sparse activation codes but also leveraging rich semantic structures and relationships between newly encountered and previously learned objects. This ability to utilize semantic similarities is crucial for efficient learning and knowledge consolidation, yet is often underutilized in current continual learning approaches. To bridge this gap, we propose Semantic-Aware Representation Learning (SARL) which employs sparse activations and a principled approach to evaluate similarities between objects encountered across different tasks and subsequently uses them to guide representation learning. Using these relationships, SARL enhances the reusability of features and reduces interference between tasks. This approach empowers the model to adapt to new information while maintaining stability, significantly improving performance in complex incremental learning scenarios. Our analysis demonstrates that SARL achieves a superior balance between plasticity and stability by harnessing the underlying semantic structure.

Abstract Top-k selection algorithms are fundamental in a wide range of applications, including high-performance computing, information retrieval, big data processing, and neural network model training. In this paper, we present RTop-K, a highly efficient parallel row-wise top-k selection algorithm specifically designed for GPUs. RTop-K leverages a binary search-based approach to optimize row-wise top-k selection, providing a scalable and accelerated solution. We conduct a detailed analysis of early stopping in our algorithm, showing that it effectively maintains the testing accuracy of neural network models while substantially improving performance. Our GPU implementation of RTop-K demonstrates superior performance over state-of-the-art row-wise top-k GPU implementations, achieving an average speed-up of up to 11.49× with early stopping and 7.29× without early stopping. Moreover, RTop-K accelerates the overall training workflow of MaxK-GNNs, delivering speed-ups ranging from 11.97% to 33.29% across different models and datasets.

Vision-Language Models (VLMs) have demonstrated impressive performance across a versatile set of tasks. A key challenge in accelerating VLMs is storing and accessing the large Key-Value (KV) cache that encodes long visual contexts, such as images or videos. While existing KV cache compression methods are effective for Large Language Models (LLMs), directly migrating them to VLMs yields suboptimal accuracy and speedup. To bridge the gap, we propose VL-Cache, a novel KV cache compression recipe tailored for accelerating VLM inference. In this paper, we first investigate the unique sparsity pattern of VLM attention by distinguishing visual and text tokens in prefill and decoding phases. Based on these observations, we introduce a layer-adaptive sparsity-aware cache budget allocation method that effectively distributes the limited cache budget across different layers, further reducing KV cache size without compromising accuracy. Additionally, we develop a modality-aware token scoring policy to better evaluate the token importance. Empirical results on multiple benchmark datasets demonstrate that retaining only 10% of KV cache achieves accuracy comparable to that with full cache. In a speed benchmark, our method accelerates end-to-end latency of generating 100 tokens by up to 2.33x and speeds up decoding by up to 7.08x, while reducing the memory footprint of KV cache in GPU by 90%.

Sparse Principal Component Analysis (Sparse PCA) is a pivotal tool in data analysis and dimensionality reduction. However, Sparse PCA is a challenging problem in both theory and practice: it is known to be NP-hard and current exact methods generally require exponential runtime. In this paper, we propose a novel framework to efficiently approximate Sparse PCA by (i) approximating the general input covariance matrix with a re-sorted block-diagonal matrix, (ii) solving the Sparse PCA sub-problem in each block, and (iii) reconstructing the solution to the original problem. Our framework is simple and powerful: it can leverage any off-the-shelf Sparse PCA algorithm and achieve significant computational speedups, with a minor additive error that is linear in the approximation error of the block-diagonal matrix. Suppose $g(k, d)$ is the runtime of an algorithm (approximately) solving Sparse PCA in dimension $d$ and with sparsity constant $k$. Our framework, when integrated with this algorithm, reduces the runtime to $\mathcal{O}\left(\frac{d}{d^\star} \cdot g(k, d^\star) + d^2\right)$, where $d^\star \leq d$ is the largest block size of the block-diagonal matrix. For instance, integrating our framework with the Branch-and-Bound algorithm reduces the complexity from $g(k, d) = \mathcal{O}(k^3\cdot d^k)$ to $\mathcal{O}(k^3\cdot d \cdot (d^\star)^{k-1})$, demonstrating exponential speedups if $d^\star$ is small. We perform large-scale evaluations on many real-world datasets: for exact Sparse PCA algorithm, our method achieves an average speedup factor of 100.50, while maintaining an average approximation error of 0.61%; for approximate Sparse PCA algorithm, our method achieves an average speedup factor of 6.00 and an average approximation error of -0.91%, meaning that our method oftentimes finds better solutions.

Neural networks are often challenging to work with due to their large size and complexity. To address this, various methods aim to reduce model size by sparsifying or decomposing weight matrices, such as magnitude pruning and low-rank or block-diagonal factorization. In this work, we present Double Sparse Factorization (DSF), where we factorize each weight matrix into two sparse matrices. Although solving this problem exactly is computationally infeasible, we propose an efficient heuristic based on alternating minimization via ADMM that achieves state-of-the-art results, enabling unprecedented sparsification of neural networks. For instance, in a one-shot pruning setting, our method can reduce the size of the LLaMA2-13B model by 50% while maintaining better performance than the dense LLaMA2-7B model. We also compare favorably with Optimal Brain Compression, the state-of-the-art layer-wise pruning approach for convolutional neural networks. Furthermore, accuracy improvements of our method persist even after further model fine-tuning. Code available at: https://github.com/usamec/double_sparse

Chain-of-thought (CoT) significantly enhances the reasoning performance of large language models (LLM). While current theoretical studies often attribute this improvement to increased expressiveness and computational capacity, we argue that expressiveness is not the primary limitation in the LLM regime, as current large models will fail on simple tasks. Using a parity-learning setup, we demonstrate that CoT can substantially improve sample efficiency even when the representation power is sufficient. Specifically, with CoT, a transformer can learn the function within polynomial samples, whereas without CoT, the required sample size is exponential. Additionally, we show that CoT simplifies the learning process by introducing sparse sequential dependencies among input tokens, and leads to a sparse and interpretable attention. We validate our theoretical analysis with both synthetic and real-world experiments, confirming that sparsity in attention layers is a key factor of the improvement induced by CoT.

Large Language Models (LLMs), while demonstrating remarkable capabilities across various applications, present significant challenges during inference due to their substantial model size, especially when deployed on edge devices. Activation sparsity offers a promising solution to reduce computation and memory movement, enabling more efficient inference, particularly for small-batch on-device applications. However, current approaches face limitations with non-ReLU activation function, which are foundational to most advanced LLMs, or require heavy continual training. Additionally, the difficulty in predicting active channels and limited achievable sparsity ratios constrain the effectiveness of activation sparsity-based methods. In this paper, we introduce R-Sparse, a training-free activation sparsity approach capable of achieving high sparsity levels in advanced LLMs. We conducted two preliminary investigations into how different components contribute to the output within a single linear layer and found two key observations: (i) the non-sparse components of the input function can be regarded as a few bias terms, and (ii) The full computation can be effectively approximated by an appropriate combination of input channels and weight singular values. Building on this, we replace the linear layers in LLMs with a rank-aware sparse inference method that leverages the sparsity of input channels and singular value components, eliminating the need for active channel prediction like the output sparsity based approaches. Experiments on Llama-2/3 and Mistral models across ten diverse tasks demonstrate that R-Sparse achieves comparable performance at 50\% model-level sparsity, resulting in a significant 43\% end-to-end efficient improvements with customized kernels.

Sparsely activated Mixture-of-Experts (MoE) models are widely adopted to scale up model capacity without increasing the computation budget. However, vanilla TopK routers are trained in a discontinuous, non-differentiable way, limiting their performance and scalability. To address this issue, we propose ReMoE, a fully differentiable MoE architecture that offers a simple yet effective drop-in replacement for the conventional TopK+Softmax routing, utilizing ReLU as the router instead. We further propose methods to regulate the router's sparsity while balancing the load among experts. ReMoE’s continuous nature enables efficient dynamic allocation of computation across tokens and layers, while also exhibiting domain specialization. Our experiments demonstrate that ReMoE consistently outperforms vanilla TopK-routed MoE across various model sizes, expert counts, and levels of granularity. Furthermore, ReMoE exhibits superior scalability with respect to the number of experts, surpassing traditional MoE architectures. The implementation based on Megatron-LM is available at https://github.com/thu-ml/ReMoE.