Data-free knowledge distillation (DFKD) has emerged as a pivotal technique in the domain of model compression, substantially reducing the dependency on the original training data. Nonetheless, conventional DFKD methods that employ synthesized training data are prone to the limitations of inadequate diversity and discrepancies in distribution between the synthesized and original datasets. To address these challenges, this paper introduces an innovative approach to DFKD through diverse diffusion augmentation (DDA). Specifically, we revise the paradigm of common data synthesis in DFKD to a composite process through leveraging diffusion models subsequent to data synthesis for self-supervised augmentation, which generates a spectrum of data samples with similar distributions while retaining controlled variations. Furthermore, to mitigate excessive deviation in the embedding space, we introduce an image filtering technique grounded in cosine similarity to maintain fidelity during the knowledge distillation process. Comprehensive experiments conducted on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets showcase the superior performance of our method across various teacher-student network configurations, outperforming the contemporary state-of-the-art DFKD methods. Code will be available at:https://github.com/SLGSP/DDA.
As large language models (LLMs) are widely applied across various fields, model compression has become increasingly crucial for reducing costs and improving inference efficiency. Post-training pruning is a promising method that does not require resource-intensive iterative training and only needs a small amount of calibration data to assess the importance of parameters. Previous research has primarily focused on designing advanced pruning methods, while different calibration data's impact on pruning performance still lacks systematical exploration. We fill this blank and surprisingly observe that the effects of calibration data even value more than designing advanced pruning strategies, especially for high sparsity. Our preliminary exploration also discloses that using calibration data similar to the training data can yield better performance. As pre-training data is usually inaccessible for advanced LLMs, we further provide a self-generating calibration data synthesis strategy to construct feasible calibration data. We conduct experiments on the recent strong open-source LLMs (e.g., DCLM, and LLaMA-3), and the results show that the proposed method outperforms commonly used calibration data and can effectively enhance strong pruning methods (e.g., Wanda, OWL).
Providing human-understandable insights into the inner workings of neural networks is an important step toward achieving more explainable and trustworthy AI. Existing approaches to such mechanistic interpretability typically require substantial prior knowledge and manual effort, with strategies tailored to specific tasks. In this work, we take a step toward automating the understanding of the network by investigating the existence of distinct sub-networks. Specifically, we explore a novel automated and task-agnostic approach based on the notion of functionally similar representations within neural networks to identify similar and dissimilar layers, revealing potential sub-networks. We achieve this by proposing, for the first time to our knowledge, the use of Gromov-Wasserstein distance, which overcomes challenges posed by varying distributions and dimensionalities across intermediate representations, issues that complicate direct layer to layer comparisons. On algebraic, language, and vision tasks, we observe the emergence of sub-groups within neural network layers corresponding to functional abstractions. Through downstream applications of model compression and fine-tuning, we show the proposed approach offers meaningful insights into the behavior of neural networks with minimal human and computational cost.
Quantization and pruning are fundamental approaches for model compression, enabling efficient inference for language models. In a post-training setting, state-of-the-art quantization and pruning methods require calibration data, a small set of unlabeled examples. Conventionally, randomly sampled web text is used, aiming to reflect the model training data. However, this poses two key problems: (1) unrepresentative calibration examples can harm model performance, and (2) organizations increasingly avoid releasing model training data. In this paper, we propose self-calibration as a solution. Our approach requires no external data, instead leveraging the model itself to generate synthetic calibration data as a better approximation of the pre-training data distribution. We extensively compare the performance of self-calibration with several baselines, across a variety of models, compression methods, and tasks. Our approach proves consistently competitive in maximizing downstream task performance, frequently outperforming even using real data.
The high computational costs of large language models (LLMs) have led to a flurry of research on LLM compression, via methods such as quantization, sparsification, or structured pruning. A new frontier in this area is given by \emph{dynamic, non-uniform} compression methods, which adjust the compression levels (e.g., sparsity) per-block or even per-layer in order to minimize accuracy loss, while guaranteeing a global compression threshold. Yet, current methods rely on heuristics for identifying the "importance" of a given layer towards the loss, based on assumptions such as \emph{error monotonicity}, i.e. that the end-to-end model compression error is proportional to the sum of layer-wise errors. In this paper, we revisit this area, and propose a new and general approach for dynamic compression that is provably optimal in a given input range. We begin from the motivating observation that, in general, \emph{error monotonicity does not hold for LLMs}: compressed models with lower sum of per-layer errors can perform \emph{worse} than models with higher error sums. To address this, we propose a new general evolutionary framework for dynamic LLM compression called EvoPress, which has provable convergence, and low sample and evaluation complexity. We show that these theoretical guarantees lead to highly competitive practical performance for dynamic compression of Llama, Mistral and Phi models. Via EvoPress, we set new state-of-the-art results across all compression approaches: structural pruning (block/layer dropping), unstructured sparsity, as well as quantization with dynamic bitwidths. Our code is available at https://github.com/IST-DASLab/EvoPress.
Knowledge distillation is a mainstream algorithm in model compression by transferring knowledge from the larger model (teacher) to the smaller model (student) to improve the performance of student. Despite many efforts, existing methods mainly investigate the consistency between instance-level feature representation or prediction, which neglects the category-level information and the difficulty of each sample, leading to undesirable performance. To address these issues, we propose a novel preview-based category contrastive learning method for knowledge distillation (PCKD). It first distills the structural knowledge of both instance-level feature correspondence and the relation between instance features and category centers in a contrastive learning fashion, which can explicitly optimize the category representation and explore the distinct correlation between representations of instances and categories, contributing to discriminative category centers and better classification results. Besides, we introduce a novel preview strategy to dynamically determine how much the student should learn from each sample according to their difficulty. Different from existing methods that treat all samples equally and curriculum learning that simply filters out hard samples, our method assigns a small weight for hard instances as a preview to better guide the student training. Extensive experiments on several challenging datasets, including CIFAR-100 and ImageNet, demonstrate the superiority over state-of-the-art methods.
Model compression methods are used to reduce the computation and energy requirements for Large Language Models (LLMs). Quantization Aware Training (QAT), an effective model compression method, is proposed to reduce performance degradation after quantization. To further minimize this degradation, we introduce two continuous approximations to the QAT process on the rounding function, traditionally approximated by the Straight-Through Estimator (STE), and the clamping function. By applying both methods, the perplexity (PPL) on the WikiText-v2 dataset of the quantized model reaches 9.0815, outperforming 9.9621 by the baseline. Also, we achieve a 2.76% improvement on BoolQ, and a 5.47% improvement on MMLU, proving that the step sizes and weights can be learned more accurately with our approach. Our method achieves better performance with the same precision, model size, and training setup, contributing to the development of more energy-efficient LLMs technology that aligns with global sustainability goals.
Convolutional neural networks (CNNs) have made significant advances in computer vision tasks, yet their high inference times and latency often limit real-world applicability. While model compression techniques have gained popularity as solutions, they often overlook the critical balance between low latency and uncompromised accuracy. By harnessing quantum-inspired pruning, tensor decomposition, and annealing-based matrix factorization - three quantum-inspired concepts - we introduce QIANets: a novel approach of redesigning the traditional GoogLeNet, DenseNet, and ResNet-18 model architectures to process more parameters and computations whilst maintaining low inference times. Despite experimental limitations, the method was tested and evaluated, demonstrating reductions in inference times, along with effective accuracy preservations.
Conventional model compression techniques for LLMs address high memory consumption and slow inference challenges but typically require computationally expensive retraining to preserve accuracy. In contrast, one-shot compression methods eliminate retraining cost, but struggle to achieve accuracy comparable to dense models. This paper presents SLIM, a new one-shot compression framework that holistically integrates hardware-friendly quantization, sparsity, and low-rank approximation into a unified process. First, we formulate the quantization process using a probabilistic approach (SLIM-Quant) that enables us to apply uniform quantization. Then, we use an existing one-shot pruning method to apply semi-structured sparsity on top of the quantized weights. Finally, to compensate for the introduced aggregated quantization and sparsity error, we use a novel saliency function with unique invertible and additive features that enables us to mathematically compute the value of low-rank adapters. SLIM improves model accuracy by up to 5.66% (LLaMA-2-7B) for 2:4 sparsity with 4-bit weight quantization, outperforming prior methods. Models compressed with SLIM achieve up to 4.3x and 3.8x on Nvidia RTX3060 and A100 GPUs, respectively. Additionally, they achieve up to 0.23x end-to-end memory reduction in comparison to their dense counterparts. We also propose an optional PEFT recipe that further improves accuracy by up to 1.66% (LLaMA-2-13B) compared to SLIM without fine-tuning.
Knowledge Distillation is a commonly used Deep Neural Network (DNN) compression method, which often maintains overall generalization performance. However, we show that even for balanced image classification datasets, such as CIFAR-100, Tiny ImageNet and ImageNet, as many as 41% of the classes are statistically significantly affected by distillation when comparing class-wise accuracy (i.e. class bias) between a teacher/distilled student or distilled student/non-distilled student model. Changes in class bias are not necessarily an undesirable outcome when considered outside of the context of a model's usage. Using two common fairness metrics, Demographic Parity Difference (DPD) and Equalized Odds Difference (EOD) on models trained with the CelebA, Trifeature, and HateXplain datasets, our results suggest that increasing the distillation temperature improves the distilled student model's fairness, and the distilled student fairness can even surpass the fairness of the teacher model at high temperatures. Additionally, we examine individual fairness, ensuring similar instances receive similar predictions. Our results confirm that higher temperatures also improve the distilled student model's individual fairness. This study highlights the uneven effects of distillation on certain classes and its potentially significant role in fairness, emphasizing that caution is warranted when using distilled models for sensitive application domains.
Post-Training Quantization (PTQ) is an effective technique for compressing Large Language Models (LLMs). While many studies focus on quantizing both weights and activations, it is still a challenge to maintain the accuracy of LLM after activating quantization. To investigate the primary cause, we extend the concept of kernel from linear algebra to quantization functions to define a new term, "quantization kernel", which refers to the set of elements in activations that are quantized to zero. Through quantitative analysis of the quantization kernel, we find that these elements are crucial for maintaining the accuracy of quantized LLMs. With the decrease of quantization kernel, the precision of quantized LLMs increases. If the quantization kernel proportion is kept below 19% for OPT models and below 1% for LLaMA models, the precision loss from quantizing activations to INT8 becomes negligible. Motivated by the goal of developing a quantization method with small quantization kernel, we propose CrossQuant: a simple yet effective method for quantizing activations. CrossQuant cross-quantizes elements using row and column-wise absolute maximum vectors, achieving a quantization kernel of approximately 16% for OPT models and less than 0.1% for LLaMA models. Experimental results on LLMs (LLaMA, OPT) ranging from 6.7B to 70B parameters demonstrate that CrossQuant improves or maintains perplexity and accuracy in language modeling, zero-shot, and few-shot tasks.
We propose ESPACE, an LLM compression technique based on dimensionality reduction of activations. Unlike prior works on weight-centric tensor decomposition, ESPACE projects activations onto a pre-calibrated set of principal components. The activation-centrality of the approach enables retraining LLMs with no loss of expressivity; while at inference, weight decomposition is obtained as a byproduct of matrix multiplication associativity. Theoretical results on the construction of projection matrices with optimal computational accuracy are provided. Experimentally, we find ESPACE enables 50% compression of GPT3, Llama2, and Nemotron4 models with small accuracy degradation, as low as a 0.18 perplexity increase on GPT3-22B. At lower compression rates of 20% to 40%, ESPACE drives GPT3 models to outperforming their baseline, by up to a 0.38 decrease in perplexity for GPT3-8B. ESPACE also reduces GEMM execution time and prefill inference latency on existing hardware. Comparison with related works on compressing Llama2-7B via matrix factorization shows that ESPACE is a first step in advancing the state-of-the-art in tensor decomposition compression of LLMs.
Large language models (LLMs) exhibit remarkable reasoning abilities, allowing them to generalize across a wide range of downstream tasks, such as commonsense reasoning or instruction following. However, as LLMs scale, inference costs become increasingly prohibitive, accumulating significantly over their life cycle. This poses the question: Can we compress pre-trained LLMs to meet diverse size and latency requirements? We leverage Neural Architecture Search (NAS) to compress LLMs by pruning structural components, such as attention heads, neurons, and layers, aiming to achieve a Pareto-optimal balance between performance and efficiency. While NAS already achieved promising results on small language models in previous work, in this paper we propose various extensions that allow us to scale to LLMs. Compared to structural pruning baselines, we show that NAS improves performance up to 3.4% on MMLU with an on-device latency speedup.
Compressive adaptation approaches, such as QLoRA, are widely popular alternatives for reducing memory requirements during fine-tuning of large language models (LLMs) while producing models capable of handling various downstream tasks. The key idea is to employ a "two-tower" architecture: compressing pre-trained LLM parameters into compact representations and fine-tuning the additive full-precision adapter, which typically has few tunable parameters in low-rank format. However, the strict algebraic assumptions, such as low-rank assumption, and the complexity of composing two-tower architectures are some of the known shortcomings, resulting in a poor accuracy-efficiency trade-off. In response to these known limitations, we propose SpaLLM (Sketched Parameter Adaptation of LLMs), a novel compressive adaptation approach for LLMs. This method is also the first to illustrate parameter-sharing compression methods for LLM fine-tuning, which, unlike QLoRA, are free from strict low-rank algebraic assumptions on adapters. Furthermore, our proposal unifies model compression and adaptation into a single, streamlined process, eliminating the need for two-tower architectures. SpaLLM sketches pre-trained LLM weights into lookup tables and directly fine-tunes the values in these tables. This approach simplifies LLMs' compressive adaptation workflow, potentially improves multi-user serving efficiency, and delivers significantly better accuracy for both natural language understanding and generation tasks. Moreover, by avoiding the "two-tower" architecture, our framework only requires one compressed matrix multiplication per layer during inference, demonstrating superior inference efficiency compared to previous methods.
Domain Generalization (DG) aims to train models that perform well not only on the training (source) domains but also on novel, unseen target data distributions. A key challenge in DG is preventing overfitting to source domains, which can be mitigated by finding flatter minima in the loss landscape. In this work, we propose Quantization-aware Training for Domain Generalization (QT-DoG) and demonstrate that weight quantization effectively leads to flatter minima in the loss landscape, thereby enhancing domain generalization. Unlike traditional quantization methods focused on model compression, QT-DoG exploits quantization as an implicit regularizer by inducing noise in model weights, guiding the optimization process toward flatter minima that are less sensitive to perturbations and overfitting. We provide both theoretical insights and empirical evidence demonstrating that quantization inherently encourages flatter minima, leading to better generalization across domains. Moreover, with the benefit of reducing the model size through quantization, we demonstrate that an ensemble of multiple quantized models further yields superior accuracy than the state-of-the-art DG approaches with no computational or memory overheads. Our extensive experiments demonstrate that QT-DoG generalizes across various datasets, architectures, and quantization algorithms, and can be combined with other DG methods, establishing its versatility and robustness.
Large Language Models (LLMs) have achieved remarkable breakthroughs. However, the huge number of parameters in LLMs require significant amount of memory storage in inference, which prevents their practical deployment in many applications. To reduce memory storage of LLMs, singular value decomposition (SVD) provides a promising solution to approximate weight matrices for compressing LLMs. In this paper, we take a step further to explore parameter sharing across different layers with SVD to achieve more effective compression for LLMs. Specifically, weight matrices in different layers are decomposed and represented as a linear combination of a set of shared basis vectors and unique coefficients. The types of weight matrices and the layer selection for basis sharing are examined when compressing LLMs to maintain the performance. Comprehensive experiments demonstrate that Basis Sharing outperforms state-of-the-art SVD-based compression approaches and parameter sharing techniques, especially under large compression ratios. Code is available at: https://github.com/TUDa-HWAI/Basis_Sharing
The disparity in healthcare personnel expertise and medical resources across different regions of the world is a pressing social issue. Artificial intelligence technology offers new opportunities to alleviate this issue. Segment Anything Model (SAM), which excels in intelligent image segmentation, has demonstrated exceptional performance in medical monitoring and assisted diagnosis. Unfortunately, the huge computational and storage overhead of SAM poses significant challenges for deployment on resource-limited edge devices. Quantization is an effective solution for model compression; however, traditional methods rely heavily on original data for calibration, which raises widespread concerns about medical data privacy and security. In this paper, we propose a data-free quantization framework for SAM, called DFQ-SAM, which learns and calibrates quantization parameters without any original data, thus effectively preserving data privacy during model compression. Specifically, we propose pseudo-positive label evolution for segmentation, combined with patch similarity, to fully leverage the semantic and distribution priors in pre-trained models, which facilitates high-quality data synthesis as a substitute for real data. Furthermore, we introduce scale reparameterization to ensure the accuracy of low-bit quantization. We perform extensive segmentation experiments on various datasets, and DFQ-SAM consistently provides significant performance on low-bit quantization. DFQ-SAM eliminates the need for data transfer in cloud-edge collaboration, thereby protecting sensitive data from potential attacks. It enables secure, fast, and personalized healthcare services at the edge, which enhances system efficiency and optimizes resource allocation, and thus facilitating the pervasive application of artificial intelligence in worldwide healthcare.
Matrix and tensor-guided parametrization for Natural Language Processing (NLP) models is fundamentally useful for the improvement of the model's systematic efficiency. However, the internal links between these two algebra structures and language model parametrization are poorly understood. Also, the existing matrix and tensor research is math-heavy and far away from machine learning (ML) and NLP research concepts. These two issues result in the recent progress on matrices and tensors for model parametrization being more like a loose collection of separate components from matrix/tensor and NLP studies, rather than a well-structured unified approach, further hindering algorithm design. To this end, we propose a unified taxonomy, which bridges the matrix/tensor compression approaches and model compression concepts in ML and NLP research. Namely, we adopt an elementary concept in linear algebra, that of a subspace, which is also the core concept in geometric algebra, to reformulate the matrix/tensor and ML/NLP concepts (e.g. attention mechanism) under one umbrella. In this way, based on our subspace formalization, typical matrix and tensor decomposition algorithms can be interpreted as geometric transformations. Finally, we revisit recent literature on matrix- or tensor-guided language model compression, rephrase and compare their core ideas, and then point out the current research gap and potential solutions.
Fluorescence lifetime imaging (FLI) is an important technique for studying cellular environments and molecular interactions, but its real-time application is limited by slow data acquisition, which requires capturing large time-resolved images and complex post-processing using iterative fitting algorithms. Deep learning (DL) models enable real-time inference, but can be computationally demanding due to complex architectures and large matrix operations. This makes DL models ill-suited for direct implementation on field-programmable gate array (FPGA)-based camera hardware. Model compression is thus crucial for practical deployment for real-time inference generation. In this work, we focus on compressing recurrent neural networks (RNNs), which are well-suited for FLI time-series data processing, to enable deployment on resource-constrained FPGA boards. We perform an empirical evaluation of various compression techniques, including weight reduction, knowledge distillation (KD), post-training quantization (PTQ), and quantization-aware training (QAT), to reduce model size and computational load while preserving inference accuracy. Our compressed RNN model, Seq2SeqLite, achieves a balance between computational efficiency and prediction accuracy, particularly at 8-bit precision. By applying KD, the model parameter size was reduced by 98\% while retaining performance, making it suitable for concurrent real-time FLI analysis on FPGA during data capture. This work represents a big step towards integrating hardware-accelerated real-time FLI analysis for fast biological processes.
Deep Multi-agent Reinforcement Learning (MARL) relies on neural networks with numerous parameters in multi-agent scenarios, often incurring substantial computational overhead. Consequently, there is an urgent need to expedite training and enable model compression in MARL. This paper proposes the utilization of dynamic sparse training (DST), a technique proven effective in deep supervised learning tasks, to alleviate the computational burdens in MARL training. However, a direct adoption of DST fails to yield satisfactory MARL agents, leading to breakdowns in value learning within deep sparse value-based MARL models. Motivated by this challenge, we introduce an innovative Multi-Agent Sparse Training (MAST) framework aimed at simultaneously enhancing the reliability of learning targets and the rationality of sample distribution to improve value learning in sparse models. Specifically, MAST incorporates the Soft Mellowmax Operator with a hybrid TD-($\lambda$) schema to establish dependable learning targets. Additionally, it employs a dual replay buffer mechanism to enhance the distribution of training samples. Building upon these aspects, MAST utilizes gradient-based topology evolution to exclusively train multiple MARL agents using sparse networks. Our comprehensive experimental investigation across various value-based MARL algorithms on multiple benchmarks demonstrates, for the first time, significant reductions in redundancy of up to $20\times$ in Floating Point Operations (FLOPs) for both training and inference, with less than $3\%$ performance degradation.