Knowledge distillation, a widely used model compression technique, works on the basis of transferring knowledge from a cumbersome teacher model to a lightweight student model. The technique involves jointly optimizing the task specific and knowledge distillation losses with a weight assigned to them. Despite these weights playing a crucial role in the performance of the distillation process, current methods provide equal weight to both losses, leading to suboptimal performance. In this paper, we propose Adaptive Knowledge Distillation, a novel technique inspired by curriculum learning to adaptively weigh the losses at instance level. This technique goes by the notion that sample difficulty increases with teacher loss. Our method follows a plug-and-play paradigm that can be applied on top of any task-specific and distillation objectives. Experiments show that our method performs better than conventional knowledge distillation method and existing instance-level loss functions.
Recent large language models (LLMs) employ billions of parameters to enable broad problem-solving capabilities. Such language models also tend to be memory-bound because of the dominance of matrix-vector and matrix-matrix multiplications with low arithmetic intensity. Therefore, optimizing the memory footprint and traffic is an important optimization direction for LLMs today. Model compression methods such as quantization and parameter pruning have been actively explored to achieve memory footprint and traffic optimization. However, the accuracy-efficiency trade-off of rank pruning (i.e., low-rank decomposition) for LLMs is not well-understood yet. Therefore, in this work, we characterize the accuracy-efficiency trade-off of a low-rank decomposition method, specifically Tucker decomposition, on recent language models, including an open-source LLM, Llama 2. We formalize the low-rank decomposition design space and show that the decomposition design space is enormous (e.g., O($2^{39}$) for Llama2-7B). To navigate such a vast design space, we formulate it and perform thorough case studies of accuracy-efficiency trade-offs using six widely used LLM benchmarks on BERT and Llama 2 models. Our results show that we can achieve a 9\% model size reduction with minimal accuracy drops, which range from 4\%p (\%p refers to "percentage point," which refers to the absolute difference between two percentage numbers; 74\% -> 78\% = 4\%p increase) to 10\%p, depending on the difficulty of the benchmark, without any retraining to recover accuracy after decomposition. The results show that low-rank decomposition can be a promising direction for LLM-based applications that require real-time service at scale (e.g., AI agent and real-time coding assistant), where the latency is as important as the model accuracy.
Deep neural networks (DNNs) have been widely used in many artificial intelligence (AI) tasks. However, deploying them brings significant challenges due to the huge cost of memory, energy, and computation. To address these challenges, researchers have developed various model compression techniques such as model quantization and model pruning. Recently, there has been a surge in research of compression methods to achieve model efficiency while retaining the performance. Furthermore, more and more works focus on customizing the DNN hardware accelerators to better leverage the model compression techniques. In addition to efficiency, preserving security and privacy is critical for deploying DNNs. However, the vast and diverse body of related works can be overwhelming. This inspires us to conduct a comprehensive survey on recent research toward the goal of high-performance, cost-efficient, and safe deployment of DNNs. Our survey first covers the mainstream model compression techniques such as model quantization, model pruning, knowledge distillation, and optimizations of non-linear operations. We then introduce recent advances in designing hardware accelerators that can adapt to efficient model compression approaches. Additionally, we discuss how homomorphic encryption can be integrated to secure DNN deployment. Finally, we discuss several issues, such as hardware evaluation, generalization, and integration of various compression approaches. Overall, we aim to provide a big picture of efficient DNNs, from algorithm to hardware accelerators and security perspectives.
Malnutrition among newborns is a top public health concern in developing countries. Identification and subsequent growth monitoring are key to successful interventions. However, this is challenging in rural communities where health systems tend to be inaccessible and under-equipped, with poor adherence to protocol. Our goal is to equip health workers and public health systems with a solution for contactless newborn anthropometry in the community. We propose NurtureNet, a multi-task model that fuses visual information (a video taken with a low-cost smartphone) with tabular inputs to regress multiple anthropometry estimates including weight, length, head circumference, and chest circumference. We show that visual proxy tasks of segmentation and keypoint prediction further improve performance. We establish the efficacy of the model through several experiments and achieve a relative error of 3.9% and mean absolute error of 114.3 g for weight estimation. Model compression to 15 MB also allows offline deployment to low-cost smartphones.
Motivated by the huge success of Transformers in the field of natural language processing (NLP), Vision Transformers (ViTs) have been rapidly developed and achieved remarkable performance in various computer vision tasks. However, their huge model sizes and intensive computations hinder ViTs' deployment on embedded devices, calling for effective model compression methods, such as quantization. Unfortunately, due to the existence of hardware-unfriendly and quantization-sensitive non-linear operations, particularly {Softmax}, it is non-trivial to completely quantize all operations in ViTs, yielding either significant accuracy drops or non-negligible hardware costs. In response to challenges associated with \textit{standard ViTs}, we focus our attention towards the quantization and acceleration for \textit{efficient ViTs}, which not only eliminate the troublesome Softmax but also integrate linear attention with low computational complexity, and propose Trio-ViT accordingly. Specifically, at the algorithm level, we develop a {tailored post-training quantization engine} taking the unique activation distributions of Softmax-free efficient ViTs into full consideration, aiming to boost quantization accuracy. Furthermore, at the hardware level, we build an accelerator dedicated to the specific Convolution-Transformer hybrid architecture of efficient ViTs, thereby enhancing hardware efficiency. Extensive experimental results consistently prove the effectiveness of our Trio-ViT framework. {Particularly, we can gain up to $\uparrow$$\mathbf{3.6}\times$, $\uparrow$$\mathbf{5.0}\times$, and $\uparrow$$\mathbf{7.3}\times$ FPS under comparable accuracy over state-of-the-art ViT accelerators, as well as $\uparrow$$\mathbf{6.0}\times$, $\uparrow$$\mathbf{1.5}\times$, and $\uparrow$$\mathbf{2.1}\times$ DSP efficiency.} Codes are available at \url{https://github.com/shihuihong214/Trio-ViT}.
Federated learning can train models without directly providing local data to the server. However, the frequent updating of the local model brings the problem of large communication overhead. Recently, scholars have achieved the communication efficiency of federated learning mainly by model compression. But they ignore two problems: 1) network state of each client changes dynamically; 2) network state among clients is not the same. The clients with poor bandwidth update local model slowly, which leads to low efficiency. To address this challenge, we propose a communication-efficient federated learning algorithm with adaptive compression under dynamic bandwidth (called AdapComFL). Concretely, each client performs bandwidth awareness and bandwidth prediction. Then, each client adaptively compresses its local model via the improved sketch mechanism based on his predicted bandwidth. Further, the server aggregates sketched models with different sizes received. To verify the effectiveness of the proposed method, the experiments are based on real bandwidth data which are collected from the network topology we build, and benchmark datasets which are obtained from open repositories. We show the performance of AdapComFL algorithm, and compare it with existing algorithms. The experimental results show that our AdapComFL achieves more efficient communication as well as competitive accuracy compared to existing algorithms.
The development of model compression is continuously motivated by the evolution of various neural network accelerators with ASIC or FPGA. On the algorithm side, the ultimate goal of quantization or pruning is accelerating the expensive DNN computations on low-power hardware. However, such a "design-and-deploy" workflow faces under-explored challenges in the current hardware-algorithm co-design community. First, although the state-of-the-art quantization algorithm can achieve low precision with negligible degradation of accuracy, the latest deep learning framework (e.g., PyTorch) can only support non-customizable 8-bit precision, data format, and parameter extraction. Secondly, the objective of quantization is to enable the computation with low-precision data. However, the current SoTA algorithm treats the quantized integer as an intermediate result, while the final output of the quantizer is the "discretized" floating-point values, ignoring the practical needs and adding additional workload to hardware designers for integer parameter extraction and layer fusion. Finally, the compression toolkits designed by the industry are constrained to their in-house product or a handful of algorithms. The limited degree of freedom in the current toolkit and the under-explored customization hinder the prototype ASIC or FPGA-based accelerator design. To resolve these challenges, we propose Torch2Chip, an open-sourced, fully customizable, and high-performance toolkit that supports user-designed compression followed by automatic model fusion and parameter extraction. Torch2Chip incorporates the hierarchical design workflow, and the user-customized compression algorithm will be directly packed into the deployment-ready format for prototype chip verification with either CNN or vision transformer (ViT). The code is available at https://github.com/SeoLabCornell/torch2chip.
The rapid advancement in Large Language Models (LLMs) has markedly enhanced the capabilities of language understanding and generation. However, the substantial model size poses hardware challenges, affecting both memory size for serving and inference latency for token generation. To address those challenges, we propose Dependency-aware Semi-structured Sparsity (DaSS), a novel method for the recent prevalent GLU-based LLMs pruning, which incorporates structural dependency into the weight magnitude-based unstructured pruning. We introduce an MLP-specific pruning metric that evaluates the importance of each weight by jointly considering its magnitude and its corresponding MLP intermediate activation norms. DaSS facilitates a balance between the adaptability offered by unstructured pruning and the structural consistency inherent in dependency-based structured pruning. Empirical evaluations on LLaMA2, Mistral, and Gemma model families demonstrate that DaSS not only outperforms both SparseGPT and Wanda in achieving hardware-friendly N:M sparsity patterns but also maintains the computational efficiency of Wanda.
Edge machine learning (ML) enables localized processing of data on devices and is underpinned by deep neural networks (DNNs). However, DNNs cannot be easily run on devices due to their substantial computing, memory and energy requirements for delivering performance that is comparable to cloud-based ML. Therefore, model compression techniques, such as pruning, have been considered. Existing pruning methods are problematic for edge ML since they: (1) Create compressed models that have limited runtime performance benefits (using unstructured pruning) or compromise the final model accuracy (using structured pruning), and (2) Require substantial compute resources and time for identifying a suitable compressed DNN model (using neural architecture search). In this paper, we explore a new avenue, referred to as Pruning-at-Initialization (PaI), using structured pruning to mitigate the above problems. We develop Reconvene, a system for rapidly generating pruned models suited for edge deployments using structured PaI. Reconvene systematically identifies and prunes DNN convolution layers that are least sensitive to structured pruning. Reconvene rapidly creates pruned DNNs within seconds that are up to 16.21x smaller and 2x faster while maintaining the same accuracy as an unstructured PaI counterpart.
Federated learning (FL) provides a promising collaborative framework to build a model from distributed clients, and this work investigates the carbon emission of the FL process. Cloud and edge servers hosting FL clients may exhibit diverse carbon footprints influenced by their geographical locations with varying power sources, offering opportunities to reduce carbon emissions by training local models with adaptive computations and communications. In this paper, we propose FedGreen, a carbon-aware FL approach to efficiently train models by adopting adaptive model sizes shared with clients based on their carbon profiles and locations using ordered dropout as a model compression technique. We theoretically analyze the trade-offs between the produced carbon emissions and the convergence accuracy, considering the carbon intensity discrepancy across countries to choose the parameters optimally. Empirical studies show that FedGreen can substantially reduce the carbon footprints of FL compared to the state-of-the-art while maintaining competitive model accuracy.
Among ML operators today, GEneralMatrix Multiplication (GEMM)-based operators are known to be key operators that build the main backbone of ML models. As their computational overhead dominates the overall execution time (e.g., 42.8% - 96.6% in our results), GEMM operators have been the prime optimization targets for fast ML inference. This led to advanced GPUs and accelerators available today, which provided significant boost in the GEMM performance compared to CPUs, aligned with the lesson from Amdahl's law. However, accelerating GEMM has significantly shifted the Amdahl's law's landscape for ML inference; due to the decreased GEMM execution time, the relative execution time of non-GEMM operators is now significant. Although the importance of non-GEMM performance is increasing, we have little knowledge about the non-GEMM performance horizon in the latest hardware platforms and models. Therefore, to guide non-GEMM-oriented optimizations, we conduct a thorough performance analysis of 17 widely adopted ML models in Hugging Face and Torchvision on workstation and data center platforms with/without GPUs. We discover that non-GEMM performance bottleneck is a considerable issue across all the platforms and models, accounting for 11.3% to 73.6% of total latency, on average. The challenge significantly aggravates when we apply quantization, which is a common model compression technique, due to the boosted GEMM performance and extra non-GEMM operators for dequantization and requantization. To provide insights into non-GEMM optimization targets, we demystify the most dominant non-GEMM operators for each model and deployment software. We also show that widely adopted optimizations such as operator fusion do not completely address the non-GEMM performance bottleneck, where non-GEMM operators still account for 15% to 48% of total latency.
Data-free knowledge distillation (DFKD) is a promising approach for addressing issues related to model compression, security privacy, and transmission restrictions. Although the existing methods exploiting DFKD have achieved inspiring achievements in coarse-grained classification, in practical applications involving fine-grained classification tasks that require more detailed distinctions between similar categories, sub-optimal results are obtained. To address this issue, we propose an approach called DFKD-FGVC that extends DFKD to fine-grained visual categorization~(FGVC) tasks. Our approach utilizes an adversarial distillation framework with attention generator, mixed high-order attention distillation, and semantic feature contrast learning. Specifically, we introduce a spatial-wise attention mechanism to the generator to synthesize fine-grained images with more details of discriminative parts. We also utilize the mixed high-order attention mechanism to capture complex interactions among parts and the subtle differences among discriminative features of the fine-grained categories, paying attention to both local features and semantic context relationships. Moreover, we leverage the teacher and student models of the distillation framework to contrast high-level semantic feature maps in the hyperspace, comparing variances of different categories. We evaluate our approach on three widely-used FGVC benchmarks (Aircraft, Cars196, and CUB200) and demonstrate its superior performance.
Vision Transformers (ViT) have marked a paradigm shift in computer vision, outperforming state-of-the-art models across diverse tasks. However, their practical deployment is hampered by high computational and memory demands. This study addresses the challenge by evaluating four primary model compression techniques: quantization, low-rank approximation, knowledge distillation, and pruning. We methodically analyze and compare the efficacy of these techniques and their combinations in optimizing ViTs for resource-constrained environments. Our comprehensive experimental evaluation demonstrates that these methods facilitate a balanced compromise between model accuracy and computational efficiency, paving the way for wider application in edge computing devices.
Recent years have seen significant efforts to adopt Artificial Intelligence (AI) in healthcare for various use cases, from computer-aided diagnosis to ICU triage. However, the size of AI models has been rapidly growing due to scaling laws and the success of foundational models, which poses an increasing challenge to leverage advanced models in practical applications. It is thus imperative to develop efficient models, especially for deploying AI solutions under resource-constrains or with time sensitivity. One potential solution is to perform model compression, a set of techniques that remove less important model components or reduce parameter precision, to reduce model computation demand. In this work, we demonstrate that model pruning, as a model compression technique, can effectively reduce inference cost for computational and digital pathology based analysis with a negligible loss of analysis performance. To this end, we develop a methodology for pruning the widely used U-Net-style architectures in biomedical imaging, with which we evaluate multiple pruning heuristics on nuclei instance segmentation and classification, and empirically demonstrate that pruning can compress models by at least 70% with a negligible drop in performance.
Recent success of pre-trained foundation vision-language models makes Open-Vocabulary Segmentation (OVS) possible. Despite the promising performance, this approach introduces heavy computational overheads for two challenges: 1) large model sizes of the backbone; 2) expensive costs during the fine-tuning. These challenges hinder this OVS strategy from being widely applicable and affordable in real-world scenarios. Although traditional methods such as model compression and efficient fine-tuning can address these challenges, they often rely on heuristics. This means that their solutions cannot be easily transferred and necessitate re-training on different models, which comes at a cost. In the context of efficient OVS, we target achieving performance that is comparable to or even better than prior OVS works based on large vision-language foundation models, by utilizing smaller models that incur lower training costs. The core strategy is to make our efficiency principled and thus seamlessly transferable from one OVS framework to others without further customization. Comprehensive experiments on diverse OVS benchmarks demonstrate our superior trade-off between segmentation accuracy and computation costs over previous works. Our code is available on https://github.com/Xujxyang/OpenTrans
Deep learning has been successfully applied to object detection from remotely sensed images. Images are typically processed on the ground rather than on-board due to the computation power of the ground system. Such offloaded processing causes delays in acquiring target mission information, which hinders its application to real-time use cases. For on-device object detection, researches have been conducted on designing efficient detectors or model compression to reduce inference latency. However, highly accurate two-stage detectors still need further exploitation for acceleration. In this paper, we propose a model simplification method for two-stage object detectors. Instead of constructing a general feature pyramid, we utilize only one feature extraction in the two-stage detector. To compensate for the accuracy drop, we apply a high pass filter to the RPN's score map. Our approach is applicable to any two-stage detector using a feature pyramid network. In the experiments with state-of-the-art two-stage detectors such as ReDet, Oriented-RCNN, and LSKNet, our method reduced computation costs upto 61.2% with the accuracy loss within 2.1% on the DOTAv1.5 dataset. Source code will be released.
In this paper, we investigate Bayesian model compression in federated learning (FL) to construct sparse models that can achieve both communication and computation efficiencies. We propose a decentralized Turbo variational Bayesian inference (D-Turbo-VBI) FL framework where we firstly propose a hierarchical sparse prior to promote a clustered sparse structure in the weight matrix. Then, by carefully integrating message passing and VBI with a decentralized turbo framework, we propose the D-Turbo-VBI algorithm which can (i) reduce both upstream and downstream communication overhead during federated training, and (ii) reduce the computational complexity during local inference. Additionally, we establish the convergence property for thr proposed D-Turbo-VBI algorithm. Simulation results show the significant gain of our proposed algorithm over the baselines in reducing communication overhead during federated training and computational complexity of final model.
Large Language Models are growing in size, and we expect them to continue to do so, as larger models train quicker. However, this increase in size will severely impact inference costs. Therefore model compression is important, to retain the performance of larger models, but with a reduced cost of running them. In this thesis we explore the methods of model compression, and we empirically demonstrate that the simple method of skipping latter attention sublayers in Transformer LLMs is an effective method of model compression, as these layers prove to be redundant, whilst also being incredibly computationally expensive. We observed a 21% speed increase in one-token generation for Llama 2 7B, whilst surprisingly and unexpectedly improving performance over several common benchmarks.
Large Language Models (LLMs) have ushered in a new era in Natural Language Processing, but their massive size demands effective compression techniques for practicality. Although numerous model compression techniques have been investigated, they typically rely on a calibration set that overlooks the multilingual context and results in significant accuracy degradation for low-resource languages. This paper introduces Multilingual Brain Surgeon (MBS), a novel calibration data sampling method for multilingual LLMs compression. MBS overcomes the English-centric limitations of existing methods by sampling calibration data from various languages proportionally to the language distribution of the model training datasets. Our experiments, conducted on the BLOOM multilingual LLM, demonstrate that MBS improves the performance of existing English-centric compression methods, especially for low-resource languages. We also uncover the dynamics of language interaction during compression, revealing that the larger the proportion of a language in the training set and the more similar the language is to the calibration language, the better performance the language retains after compression. In conclusion, MBS presents an innovative approach to compressing multilingual LLMs, addressing the performance disparities and improving the language inclusivity of existing compression techniques.
Knowledge distillation (KD) has become a widely used technique in the field of model compression, which aims to transfer knowledge from a large teacher model to a lightweight student model for efficient network development. In addition to the supervision of ground truth, the vanilla KD method regards the predictions of the teacher as soft labels to supervise the training of the student model. Based on vanilla KD, various approaches have been developed to further improve the performance of the student model. However, few of these previous methods have considered the reliability of the supervision from teacher models. Supervision from erroneous predictions may mislead the training of the student model. This paper therefore proposes to tackle this problem from two aspects: Label Revision to rectify the incorrect supervision and Data Selection to select appropriate samples for distillation to reduce the impact of erroneous supervision. In the former, we propose to rectify the teacher's inaccurate predictions using the ground truth. In the latter, we introduce a data selection technique to choose suitable training samples to be supervised by the teacher, thereby reducing the impact of incorrect predictions to some extent. Experiment results demonstrate the effectiveness of our proposed method, and show that our method can be combined with other distillation approaches, improving their performance.