Low-rank adaption (LoRA) is a prominent method that adds a small number of learnable parameters to the frozen pre-trained weights for parameter-efficient fine-tuning. Prompted by the question, ``Can we make its representation enough with LoRA weights solely at the final phase of finetuning without the pre-trained weights?'' In this work, we introduce Progressive Compression LoRA~(PC-LoRA), which utilizes low-rank adaptation (LoRA) to simultaneously perform model compression and fine-tuning. The PC-LoRA method gradually removes the pre-trained weights during the training process, eventually leaving only the low-rank adapters in the end. Thus, these low-rank adapters replace the whole pre-trained weights, achieving the goals of compression and fine-tuning at the same time. Empirical analysis across various models demonstrates that PC-LoRA achieves parameter and FLOPs compression rates of 94.36%/89.1% for vision models, e.g., ViT-B, and 93.42%/84.2% parameters and FLOPs compressions for language models, e.g., BERT.
This work explores knowledge distillation (KD) for visually-rich document (VRD) applications such as document layout analysis (DLA) and document image classification (DIC). While VRD research is dependent on increasingly sophisticated and cumbersome models, the field has neglected to study efficiency via model compression. Here, we design a KD experimentation methodology for more lean, performant models on document understanding (DU) tasks that are integral within larger task pipelines. We carefully selected KD strategies (response-based, feature-based) for distilling knowledge to and from backbones with different architectures (ResNet, ViT, DiT) and capacities (base, small, tiny). We study what affects the teacher-student knowledge gap and find that some methods (tuned vanilla KD, MSE, SimKD with an apt projector) can consistently outperform supervised student training. Furthermore, we design downstream task setups to evaluate covariate shift and the robustness of distilled DLA models on zero-shot layout-aware document visual question answering (DocVQA). DLA-KD experiments result in a large mAP knowledge gap, which unpredictably translates to downstream robustness, accentuating the need to further explore how to efficiently obtain more semantic document layout awareness.
Large Language Models~(LLMs) have become foundational in the realm of natural language processing, demonstrating performance improvements as model sizes increase. The Mixture-of-Experts~(MoE) approach offers a promising way to scale LLMs more efficiently by using fewer computational FLOPs through sparse activation. However, it suffers from significant memory overheads, necessitating model compression techniques. Post-training quantization, a popular method for model compression, proves less effective when directly applied to MoE models due to MoE's overlooked inherent sparsity. This paper explores several MoE structure-aware quantization heuristics, ranging from coarse to fine granularity, from MoE block to individual linear weight. Our investigations reveal critical principles: different MoE structures (i.e., blocks, experts, linear layers) require varying numbers of weight bits for effective and efficient quantization. Conclusions are supported by extensive benchmarking across two representative MoE models and six tasks. We further introduce novel enhancements to more accurately identify the most critical weights in MoE quantization that necessitate higher bit allocations, including the linear weight outlier scorer and MoE block scorer. Additionally, subsequent experiments validate our findings in the context of both weight and activation quantization.
Self-supervised learning (SSL) speech models have achieved remarkable performance in various tasks, yet the biased outcomes, especially affecting marginalized groups, raise significant concerns. Social bias refers to the phenomenon where algorithms potentially amplify disparate properties between social groups present in the data used for training. Bias in SSL models can perpetuate injustice by automating discriminatory patterns and reinforcing inequitable systems. This work reveals that prevalent SSL models inadvertently acquire biased associations. We probe how various factors, such as model architecture, size, and training methodologies, influence the propagation of social bias within these models. Finally, we explore the efficacy of debiasing SSL models through regularization techniques, specifically via model compression. Our findings reveal that employing techniques such as row-pruning and training wider, shallower models can effectively mitigate social bias within SSL model.
The ability of machine learning (ML) algorithms to generalize well to unseen data has been studied through the lens of information theory, by bounding the generalization error with the input-output mutual information (MI), i.e., the MI between the training data and the learned hypothesis. Yet, these bounds have limited practicality for modern ML applications (e.g., deep learning), due to the difficulty of evaluating MI in high dimensions. Motivated by recent findings on the compressibility of neural networks, we consider algorithms that operate by slicing the parameter space, i.e., trained on random lower-dimensional subspaces. We introduce new, tighter information-theoretic generalization bounds tailored for such algorithms, demonstrating that slicing improves generalization. Our bounds offer significant computational and statistical advantages over standard MI bounds, as they rely on scalable alternative measures of dependence, i.e., disintegrated mutual information and $k$-sliced mutual information. Then, we extend our analysis to algorithms whose parameters do not need to exactly lie on random subspaces, by leveraging rate-distortion theory. This strategy yields generalization bounds that incorporate a distortion term measuring model compressibility under slicing, thereby tightening existing bounds without compromising performance or requiring model compression. Building on this, we propose a regularization scheme enabling practitioners to control generalization through compressibility. Finally, we empirically validate our results and achieve the computation of non-vacuous information-theoretic generalization bounds for neural networks, a task that was previously out of reach.
Pre-trained large language models (LLMs) based on Transformer have demonstrated striking in-context learning (ICL) abilities. With a few demonstration input-label pairs, they can predict the label for an unseen input without any parameter updates. In this paper, we show an exciting phenomenon that SVD-based weight pruning can enhance ICL performance, and more surprising, pruning weights in deep layers often results in more stable performance improvements than in shallow layers. However, the underlying mechanism of those findings still remains an open question. To reveal those findings, we conduct an in-depth theoretical analysis by presenting the implicit gradient descent (GD) trajectories of ICL and giving the mutual information based generalization bounds of ICL via full implicit GD trajectories. This helps us reasonably explain the surprising experimental findings. Besides, based on all our experimental and theoretical insights, we intuitively propose a simple, model-compression and derivative-free algorithm for downstream tasks in enhancing ICL inference. Experiments on benchmark datasets and open source LLMs display the method effectiveness\footnote{The code is available at \url{https://github.com/chen123CtrlS/EnhancingICL_SVDPruning}.}.
Benefiting from well-trained deep neural networks (DNNs), model compression have captured special attention for computing resource limited equipment, especially edge devices. Knowledge distillation (KD) is one of the widely used compression techniques for edge deployment, by obtaining a lightweight student model from a well-trained teacher model released on public platforms. However, it has been empirically noticed that the backdoor in the teacher model will be transferred to the student model during the process of KD. Although numerous KD methods have been proposed, most of them focus on the distillation of a high-performing student model without robustness consideration. Besides, some research adopts KD techniques as effective backdoor mitigation tools, but they fail to perform model compression at the same time. Consequently, it is still an open problem to well achieve two objectives of robust KD, i.e., student model's performance and backdoor mitigation. To address these issues, we propose RobustKD, a robust knowledge distillation that compresses the model while mitigating backdoor based on feature variance. Specifically, RobustKD distinguishes the previous works in three key aspects: (1) effectiveness: by distilling the feature map of the teacher model after detoxification, the main task performance of the student model is comparable to that of the teacher model; (2) robustness: by reducing the characteristic variance between the teacher model and the student model, it mitigates the backdoor of the student model under backdoored teacher model scenario; (3) generic: RobustKD still has good performance in the face of multiple data models (e.g., WRN 28-4, Pyramid-200) and diverse DNNs (e.g., ResNet50, MobileNet).
Weighted low rank approximation (WLRA) is an important yet computationally challenging primitive with applications ranging from statistical analysis, model compression, and signal processing. To cope with the NP-hardness of this problem, prior work considers heuristics, bicriteria, or fixed parameter tractable algorithms to solve this problem. In this work, we introduce a new relaxed solution to WLRA which outputs a matrix that is not necessarily low rank, but can be stored using very few parameters and gives provable approximation guarantees when the weight matrix has low rank. Our central idea is to use the weight matrix itself to reweight a low rank solution, which gives an extremely simple algorithm with remarkable empirical performance in applications to model compression and on synthetic datasets. Our algorithm also gives nearly optimal communication complexity bounds for a natural distributed problem associated with this problem, for which we show matching communication lower bounds. Together, our communication complexity bounds show that the rank of the weight matrix provably parameterizes the communication complexity of WLRA. We also obtain the first relative error guarantees for feature selection with a weighted objective.
The emergence of deep and large-scale spiking neural networks (SNNs) exhibiting high performance across diverse complex datasets has led to a need for compressing network models due to the presence of a significant number of redundant structural units, aiming to more effectively leverage their low-power consumption and biological interpretability advantages. Currently, most model compression techniques for SNNs are based on unstructured pruning of individual connections, which requires specific hardware support. Hence, we propose a structured pruning approach based on the activity levels of convolutional kernels named Spiking Channel Activity-based (SCA) network pruning framework. Inspired by synaptic plasticity mechanisms, our method dynamically adjusts the network's structure by pruning and regenerating convolutional kernels during training, enhancing the model's adaptation to the current target task. While maintaining model performance, this approach refines the network architecture, ultimately reducing computational load and accelerating the inference process. This indicates that structured dynamic sparse learning methods can better facilitate the application of deep SNNs in low-power and high-efficiency scenarios.
Large language models~(LLMs) have recently demonstrated promising performance in many tasks. However, the high storage and computational cost of LLMs has become a challenge for deploying LLMs. Weight quantization has been widely used for model compression, which can reduce both storage and computational cost. Most existing weight quantization methods for LLMs use a rank-one codebook for quantization, which results in substantial accuracy loss when the compression ratio is high. In this paper, we propose a novel weight quantization method, called low-rank codebook based quantization~(LCQ), for LLMs. LCQ adopts a low-rank codebook, the rank of which can be larger than one, for quantization. Experiments show that LCQ can achieve better accuracy than existing methods with a negligibly extra storage cost.
The increasing size of deep neural networks (DNNs) necessitates effective model compression to reduce their computational and memory footprints. Sparsity and quantization are two prominent compression methods that have been shown to reduce DNNs' computational and memory footprints significantly while preserving model accuracy. However, how these two methods interact when combined together remains a key question for developers, as many tacitly assume that they are orthogonal, meaning that their combined use does not introduce additional errors beyond those introduced by each method independently. In this paper, we provide the first mathematical proof that sparsity and quantization are non-orthogonal. We corroborate these results with experiments spanning a range of large language models, including the OPT and LLaMA model families (with 125M to 8B parameters), and vision models like ViT and ResNet. We show that the order in which we apply these methods matters because applying quantization before sparsity may disrupt the relative importance of tensor elements, which may inadvertently remove significant elements from a tensor. More importantly, we show that even if applied in the correct order, the compounded errors from sparsity and quantization can significantly harm accuracy. Our findings extend to the efficient deployment of large models in resource-constrained compute platforms to reduce serving cost, offering insights into best practices for applying these compression methods to maximize hardware resource efficiency without compromising accuracy.
Deep learning neural network models must be large enough to adapt to their problem domain, while small enough to avoid overfitting training data during gradient descent. To balance these competing demands, overprovisioned deep learning models such as transformers are trained for a single epoch on large data sets, and hence inefficient with both computing resources and training data. In response to these inefficiencies, we exploit learning theory to derive Occam Gradient Descent, an algorithm that interleaves adaptive reduction of model size to minimize generalization error, with gradient descent on model weights to minimize fitting error. In contrast, traditional gradient descent greedily minimizes fitting error without regard to generalization error. Our algorithm simultaneously descends the space of weights and topological size of any neural network without modification. With respect to loss, compute and model size, our experiments show (a) on image classification benchmarks, linear and convolutional neural networks trained with Occam Gradient Descent outperform traditional gradient descent with or without post-train pruning; (b) on a range of tabular data classification tasks, neural networks trained with Occam Gradient Descent outperform traditional gradient descent, as well as Random Forests; (c) on natural language transformers, Occam Gradient Descent outperforms traditional gradient descent.
Federated Learning (FL) is commonly used in systems with distributed and heterogeneous devices with access to varying amounts of data and diverse computing and storage capacities. FL training process enables such devices to update the weights of a shared model locally using their local data and then a trusted central server combines all of those models to generate a global model. In this way, a global model is generated while the data remains local to devices to preserve privacy. However, training large models such as Deep Neural Networks (DNNs) on resource-constrained devices can take a prohibitively long time and consume a large amount of energy. In the current process, the low-capacity devices are excluded from the training process, although they might have access to unseen data. To overcome this challenge, we propose a model compression approach that enables heterogeneous devices with varying computing capacities to participate in the FL process. In our approach, the server shares a dense model with all devices to train it: Afterwards, the trained model is gradually compressed to obtain submodels with varying levels of sparsity to be used as suitable initial global models for resource-constrained devices that were not capable of train the first dense model. This results in an increased participation rate of resource-constrained devices while the transferred weights from the previous round of training are preserved. Our validation experiments show that despite reaching about 50 per cent global sparsity, generated submodels maintain their accuracy while can be shared to increase participation by around 50 per cent.
Although deep convolutional neural networks have achieved rapid development, it is challenging to widely promote and apply these models on low-power devices, due to computational and storage limitations. To address this issue, researchers have proposed techniques such as model compression, activation sparsity induction, and hardware accelerators. This paper presents a method to induce the sparsity of activation maps based on Transformed $\ell1$ regularization, so as to improve the research in the field of activation sparsity induction. Further, the method is innovatively combined with traditional pruning, constituting a dual sparse training framework. Compared to previous methods, Transformed $\ell1$ can achieve higher sparsity and better adapt to different network structures. Experimental results show that the method achieves improvements by more than 20\% in activation map sparsity on most models and corresponding datasets without compromising the accuracy. Specifically, it achieves a 27.52\% improvement for ResNet18 on the ImageNet dataset, and a 44.04\% improvement for LeNet5 on the MNIST dataset. In addition, the dual sparse training framework can greatly reduce the computational load and provide potential for reducing the required storage during runtime. Specifically, the ResNet18 and ResNet50 models obtained by the dual sparse training framework respectively reduce 81.7\% and 84.13\% of multiplicative floating-point operations, while maintaining accuracy and a low pruning rate.
Federated learning (FL), as an emerging collaborative learning paradigm, has garnered significant attention due to its capacity to preserve privacy within distributed learning systems. In these systems, clients collaboratively train a unified neural network model using their local datasets and share model parameters rather than raw data, enhancing privacy. Predominantly, FL systems are designed for mobile and edge computing environments where training typically occurs over wireless networks. Consequently, as model sizes increase, the conventional FL frameworks increasingly consume substantial communication resources. To address this challenge and improve communication efficiency, this paper introduces a novel hierarchical FL framework that integrates the benefits of clustered FL and model compression. We present an adaptive clustering algorithm that identifies a core client and dynamically organizes clients into clusters. Furthermore, to enhance transmission efficiency, each core client implements a local aggregation with compression (LC aggregation) algorithm after collecting compressed models from other clients within the same cluster. Simulation results affirm that our proposed algorithms not only maintain comparable predictive accuracy but also significantly reduce energy consumption relative to existing FL mechanisms.
Deep neural networks (DNNs) have heavily relied on traditional computational units like CPUs and GPUs. However, this conventional approach brings significant computational burdens, latency issues, and high power consumption, limiting their effectiveness. This has sparked the need for lightweight networks like ExtremeC3Net. On the other hand, there have been notable advancements in optical computational units, particularly with metamaterials, offering the exciting prospect of energy-efficient neural networks operating at the speed of light. Yet, the digital design of metamaterial neural networks (MNNs) faces challenges such as precision, noise, and bandwidth, limiting their application to intuitive tasks and low-resolution images. In this paper, we propose a large kernel lightweight segmentation model, ExtremeMETA. Based on the ExtremeC3Net, the ExtremeMETA maximizes the ability of the first convolution layer by exploring a larger convolution kernel and multiple processing paths. With the proposed large kernel convolution model, we extend the optic neural network application boundary to the segmentation task. To further lighten the computation burden of the digital processing part, a set of model compression methods is applied to improve model efficiency in the inference stage. The experimental results on three publicly available datasets demonstrate that the optimized efficient design improved segmentation performance from 92.45 to 95.97 on mIoU while reducing computational FLOPs from 461.07 MMacs to 166.03 MMacs. The proposed the large kernel lightweight model ExtremeMETA showcases the hybrid design's ability on complex tasks.
Decoder-only LLM-based embedding models are beginning to outperform BERT or T5-based embedding models in general-purpose text embedding tasks, including dense vector-based retrieval. In this work, we introduce NV-Embed, incorporating architectural designs, training procedures, and curated datasets to significantly enhance the performance of LLM as a versatile embedding model, while maintaining its simplicity and reproducibility. For model architecture, we propose a latent attention layer to obtain pooled embeddings, which consistently improves retrieval and downstream task accuracy compared to mean pooling or using the last <EOS> token embedding from LLMs. To enhance representation learning, we remove the causal attention mask of LLMs during contrastive training. For training algorithm, we introduce a two-stage contrastive instruction-tuning method. It first applies contrastive training with instructions on retrieval datasets, utilizing in-batch negatives and curated hard negative examples. At stage-2, it blends various non-retrieval into instruction tuning, which not only enhances non-retrieval task accuracy but also improves retrieval performance. For training data, we utilize the hard-negative mining, synthetic data generation and existing public available datasets to boost the performance of embedding model. By combining these techniques, our NV-Embed-v1 and NV-Embed-v2 models obtained the No.1 position on the MTEB leaderboard (as of May 24 and August 30, 2024, respectively) across 56 tasks, demonstrating the sustained effectiveness of the proposed methods over time. It also achieved the highest scores in the Long Doc section and the second-highest scores in the QA section of the AIR Benchmark, which covers a range of out-of-domain information retrieval topics beyond those in MTEB. We further provide the analysis of model compression techniques for generalist embedding models.
The advancement of sophisticated artificial intelligence (AI) algorithms has led to a notable increase in energy usage and carbon dioxide emissions, intensifying concerns about climate change. This growing problem has brought the environmental sustainability of AI technologies to the forefront, especially as they expand across various sectors. In response to these challenges, there is an urgent need for the development of sustainable AI solutions. These solutions must focus on energy-efficient embedded systems that are capable of handling diverse data types even in environments with limited resources, thereby ensuring both technological progress and environmental responsibility. Integrating complementary multimodal data into tiny machine learning models for edge devices is challenging due to increased complexity, latency, and power consumption. This work introduces TinyM$^2$Net-V3, a system that processes different modalities of complementary data, designs deep neural network (DNN) models, and employs model compression techniques including knowledge distillation and low bit-width quantization with memory-aware considerations to fit models within lower memory hierarchy levels, reducing latency and enhancing energy efficiency on resource-constrained devices. We evaluated TinyM$^2$Net-V3 in two multimodal case studies: COVID-19 detection using cough, speech, and breathing audios, and pose classification from depth and thermal images. With tiny inference models (6 KB and 58 KB), we achieved 92.95% and 90.7% accuracies, respectively. Our tiny machine learning models, deployed on resource limited hardware, demonstrated low latencies within milliseconds and very high power efficiency.
The internal structure and operation mechanism of large-scale language models are analyzed theoretically, especially how Transformer and its derivative architectures can restrict computing efficiency while capturing long-term dependencies. Further, we dig deep into the efficiency bottleneck of the training phase, and evaluate in detail the contribution of adaptive optimization algorithms (such as AdamW), massively parallel computing techniques, and mixed precision training strategies to accelerate convergence and reduce memory footprint. By analyzing the mathematical principles and implementation details of these algorithms, we reveal how they effectively improve training efficiency in practice. In terms of model deployment and inference optimization, this paper systematically reviews the latest advances in model compression techniques, focusing on strategies such as quantification, pruning, and knowledge distillation. By comparing the theoretical frameworks of these techniques and their effects in different application scenarios, we demonstrate their ability to significantly reduce model size and inference delay while maintaining model prediction accuracy. In addition, this paper critically examines the limitations of current efficiency optimization methods, such as the increased risk of overfitting, the control of performance loss after compression, and the problem of algorithm generality, and proposes some prospects for future research. In conclusion, this study provides a comprehensive theoretical framework for understanding the efficiency optimization of large-scale language models.
Continual learning, involving sequential training on diverse tasks, often faces catastrophic forgetting. While knowledge distillation-based approaches exhibit notable success in preventing forgetting, we pinpoint a limitation in their ability to distill the cumulative knowledge of all the previous tasks. To remedy this, we propose Dense Knowledge Distillation (DKD). DKD uses a task pool to track the model's capabilities. It partitions the output logits of the model into dense groups, each corresponding to a task in the task pool. It then distills all tasks' knowledge using all groups. However, using all the groups can be computationally expensive, we also suggest random group selection in each optimization step. Moreover, we propose an adaptive weighting scheme, which balances the learning of new classes and the retention of old classes, based on the count and similarity of the classes. Our DKD outperforms recent state-of-the-art baselines across diverse benchmarks and scenarios. Empirical analysis underscores DKD's ability to enhance model stability, promote flatter minima for improved generalization, and remains robust across various memory budgets and task orders. Moreover, it seamlessly integrates with other CL methods to boost performance and proves versatile in offline scenarios like model compression.