Large Language Models (LLMs), such as LLaMA and T5, have shown exceptional performance across various tasks through fine-tuning. Although low-rank adaption (LoRA) has emerged to cheaply fine-tune these LLMs on downstream tasks, their deployment is still hindered by the vast model scale and computational costs. Post-training model pruning offers a way to compress LLMs. However, the current pruning methods designed for LLMs are not compatible with LoRA. This is due to their utilization of unstructured pruning on LLMs, impeding the merging of LoRA weights, or their dependence on the gradients of pre-trained weights to guide pruning, which can impose significant memory overhead. To this end, we propose LoRAPrune, a new framework that delivers an accurate structured pruned model in a highly memory-efficient manner. Specifically, we first design a LoRA-guided pruning criterion, which uses the weights and gradients of LoRA, rather than the gradients of pre-trained weights for importance estimation. We subsequently integrate this criterion into an iterative pruning process, effectively removing redundant channels and heads. Extensive experimental results demonstrate the superior performance of our LoRAPrune over existing approaches on the LLaMA series models. At a 50\% compression rate, LoRAPrune demonstrates superior performance over LLM-Pruner, achieving a reduction in perplexity by 4.81 on WikiText2 and 3.46 on PTB, while also decreasing memory usage by 52.6%. Besides, LoRAPrune also matches semi-structural pruning across multiple LLMs, proving its wide applicability. The code is available at https://github.com/aim-uofa/LoRAPrune.
Large-scale pre-trained text-image models with dual-encoder architectures (such as CLIP) are typically adopted for various vision-language applications, including text-image retrieval. However,these models are still less practical on edge devices or for real-time situations, due to the substantial indexing and inference time and the large consumption of computational resources. Although knowledge distillation techniques have been widely utilized for uni-modal model compression, how to expand them to the situation when the numbers of modalities and teachers/students are doubled has been rarely studied. In this paper, we conduct comprehensive experiments on this topic and propose the fully-Connected knowledge interaction graph (Cona) technique for cross-modal pre-training distillation. Based on our findings, the resulting ConaCLIP achieves SOTA performances on the widely-used Flickr30K and MSCOCO benchmarks under the lightweight setting. An industry application of our method on an e-commercial platform further demonstrates the significant effectiveness of ConaCLIP.
Attention-based vision models, such as Vision Transformer (ViT) and its variants, have shown promising performance in various computer vision tasks. However, these emerging architectures suffer from large model sizes and high computational costs, calling for efficient model compression solutions. To date, pruning ViTs has been well studied, while other compression strategies that have been widely applied in CNN compression, e.g., model factorization, is little explored in the context of ViT compression. This paper explores an efficient method for compressing vision transformers to enrich the toolset for obtaining compact attention-based vision models. Based on the new insight on the multi-head attention layer, we develop a highly efficient ViT compression solution, which outperforms the state-of-the-art pruning methods. For compressing DeiT-small and DeiT-base models on ImageNet, our proposed approach can achieve 0.45% and 0.76% higher top-1 accuracy even with fewer parameters. Our finding can also be applied to improve the customization efficiency of text-to-image diffusion models, with much faster training (up to $2.6\times$ speedup) and lower extra storage cost (up to $1927.5\times$ reduction) than the existing works.
Large speech models are rapidly gaining traction in research community. As a result, model compression has become an important topic, so that these models can fit in memory and be served with reduced cost. Practical approaches for compressing automatic speech recognition (ASR) model use int8 or int4 weight quantization. In this study, we propose to develop 2-bit ASR models. We explore the impact of symmetric and asymmetric quantization combined with sub-channel quantization and clipping on both LibriSpeech dataset and large-scale training data. We obtain a lossless 2-bit Conformer model with 32% model size reduction when compared to state of the art 4-bit Conformer model for LibriSpeech. With the large-scale training data, we obtain a 2-bit Conformer model with over 40% model size reduction against the 4-bit version at the cost of 17% relative word error rate degradation
With the rapid increase in the size of neural networks, model compression has become an important area of research. Quantization is an effective technique at decreasing the model size, memory access, and compute load of large models. Despite recent advances in quantization aware training (QAT) technique, most papers present evaluations that are focused on computer vision tasks, which have different training dynamics compared to sequence tasks. In this paper, we first benchmark the impact of popular techniques such as straight through estimator, pseudo-quantization noise, learnable scale parameter, clipping, etc. on 4-bit seq2seq models across a suite of speech recognition datasets ranging from 1,000 hours to 1 million hours, as well as one machine translation dataset to illustrate its applicability outside of speech. Through the experiments, we report that noise based QAT suffers when there is insufficient regularization signal flowing back to the quantization scale. We propose low complexity changes to the QAT process to improve model accuracy (outperforming popular learnable scale and clipping methods). With the improved accuracy, it opens up the possibility to exploit some of the other benefits of noise based QAT: 1) training a single model that performs well in mixed precision mode and 2) improved generalization on long form speech recognition.
Multilingual language model (LM) have become a powerful tool in NLP especially for non-English languages. Nevertheless, model parameters of multilingual LMs remain large due to the larger embedding matrix of the vocabulary covering tokens in different languages. On the contrary, monolingual LMs can be trained in a target language with the language-specific vocabulary only, but this requires a large budget and availability of reliable corpora to achieve a high-quality LM from scratch. In this paper, we propose vocabulary-trimming (VT), a method to reduce a multilingual LM vocabulary to a target language by deleting irrelevant tokens from its vocabulary. In theory, VT can compress any existing multilingual LM to build monolingual LMs in any language covered by the multilingual LM. In our experiments, we show that VT can retain the original performance of the multilingual LM, while being smaller in size (in general around 50% of the original vocabulary size is enough) than the original multilingual LM. The evaluation is performed over four NLP tasks (two generative and two classification tasks) among four widely used multilingual LMs in seven languages. Finally, we show that this methodology can keep the best of both monolingual and multilingual worlds by keeping a small size as monolingual models without the need for specifically retraining them, and even limiting potentially harmful social biases.
As language models increase in size by the day, methods for efficient inference are critical to leveraging their capabilities for various applications. Prior work has investigated techniques like model pruning, knowledge distillation, and data multiplexing to increase model throughput without sacrificing accuracy. In this paper, we combine two such methods -- structured pruning and data multiplexing -- to compound the speedup gains obtained by either method. Our approach, PruMUX, obtains up to 7.5-29.5X throughput improvement over BERT-base model with accuracy threshold from 80% to 74%. We further study various combinations of parameters (such as sparsity and multiplexing factor) in the two techniques to provide a comprehensive analysis of the tradeoff between accuracy and throughput in the resulting models. We then propose Auto-PruMUX, a meta-level model that can predict the high-performance parameters for pruning and multiplexing given a desired accuracy loss budget, providing a practical method to leverage the combination effectively.
Text prediction models, when used in applications like email clients or word processors, must protect user data privacy and adhere to model size constraints. These constraints are crucial to meet memory and inference time requirements, as well as to reduce inference costs. Building small, fast, and private domain-specific language models is a thriving area of research. In this work, we show that a careful pre-training on a \emph{subset} of the public dataset that is guided by the private dataset is crucial to train small language models with differential privacy. On standard benchmarks, small models trained with our new framework achieve state-of-the-art performance. In addition to performance improvements, our results demonstrate that smaller models, through careful pre-training and private fine-tuning, can match the performance of much larger models that do not have access to private data. This underscores the potential of private learning for model compression and enhanced efficiency.
The excellent performance of deep neural networks is usually accompanied by a large number of parameters and computations, which have limited their usage on the resource-limited edge devices. To address this issue, abundant methods such as pruning, quantization and knowledge distillation have been proposed to compress neural networks and achieved significant breakthroughs. However, most of these compression methods focus on the architecture or the training method of neural networks but ignore the influence from data augmentation. In this paper, we revisit the usage of data augmentation in model compression and give a comprehensive study on the relation between model sizes and their optimal data augmentation policy. To sum up, we mainly have the following three observations: (A) Models in different sizes prefer data augmentation with different magnitudes. Hence, in iterative pruning, data augmentation with varying magnitudes leads to better performance than data augmentation with a consistent magnitude. (B) Data augmentation with a high magnitude may significantly improve the performance of large models but harm the performance of small models. Fortunately, small models can still benefit from strong data augmentations by firstly learning them with "additional parameters" and then discard these "additional parameters" during inference. (C) The prediction of a pre-trained large model can be utilized to measure the difficulty of data augmentation. Thus it can be utilized as a criterion to design better data augmentation policies. We hope this paper may promote more research on the usage of data augmentation in model compression.
While the numerous parameters in Large Language Models (LLMs) contribute to their superior performance, this massive scale makes them inefficient and memory-hungry. Thus, they are hard to deploy on commodity hardware, such as one single GPU. Given the memory and power constraints of such devices, model compression methods are widely employed to reduce both the model size and inference latency, which essentially trades off model quality in return for improved efficiency. Thus, optimizing this accuracy-efficiency trade-off is crucial for the LLM deployment on commodity hardware. In this paper, we introduce a new perspective to optimize this trade-off by prompting compressed models. Specifically, we first observe that for certain questions, the generation quality of a compressed LLM can be significantly improved by adding carefully designed hard prompts, though this isn't the case for all questions. Based on this observation, we propose a soft prompt learning method where we expose the compressed model to the prompt learning process, aiming to enhance the performance of prompts. Our experimental analysis suggests our soft prompt strategy greatly improves the performance of the 8x compressed LLaMA-7B model (with a joint 4-bit quantization and 50% weight pruning compression), allowing them to match their uncompressed counterparts on popular benchmarks. Also, we demonstrate that these learned prompts can be transferred across various datasets, tasks, and compression levels. Hence with this transferability, we can stitch the soft prompt to a newly compressed model to improve the test-time accuracy in an ``in-situ'' way.
Knowledge distillation has attracted a great deal of interest recently to compress pre-trained language models. However, existing knowledge distillation methods suffer from two limitations. First, the student model simply imitates the teacher's behavior while ignoring the underlying reasoning. Second, these methods usually focus on the transfer of sophisticated model-specific knowledge but overlook data-specific knowledge. In this paper, we present a novel attribution-driven knowledge distillation approach, which explores the token-level rationale behind the teacher model based on Integrated Gradients (IG) and transfers attribution knowledge to the student model. To enhance the knowledge transfer of model reasoning and generalization, we further explore multi-view attribution distillation on all potential decisions of the teacher. Comprehensive experiments are conducted with BERT on the GLUE benchmark. The experimental results demonstrate the superior performance of our approach to several state-of-the-art methods.
Vision Transformer (ViT) has performed remarkably in various computer vision tasks. Nonetheless, affected by the massive amount of parameters, ViT usually suffers from serious overfitting problems with a relatively limited number of training samples. In addition, ViT generally demands heavy computing resources, which limit its deployment on resource-constrained devices. As a type of model-compression method, model binarization is potentially a good choice to solve the above problems. Compared with the full-precision one, the model with the binarization method replaces complex tensor multiplication with simple bit-wise binary operations and represents full-precision model parameters and activations with only 1-bit ones, which potentially solves the problem of model size and computational complexity, respectively. In this paper, we investigate a binarized ViT model. Empirically, we observe that the existing binarization technology designed for Convolutional Neural Networks (CNN) cannot migrate well to a ViT's binarization task. We also find that the decline of the accuracy of the binary ViT model is mainly due to the information loss of the Attention module and the Value vector. Therefore, we propose a novel model binarization technique, called Group Superposition Binarization (GSB), to deal with these issues. Furthermore, in order to further improve the performance of the binarization model, we have investigated the gradient calculation procedure in the binarization process and derived more proper gradient calculation equations for GSB to reduce the influence of gradient mismatch. Then, the knowledge distillation technique is introduced to alleviate the performance degradation caused by model binarization. Analytically, model binarization can limit the parameters search space during parameter updates while training a model....
Knowledge distillation (KD) is a promising technique for model compression in neural machine translation. However, where the knowledge hides in KD is still not clear, which may hinder the development of KD. In this work, we first unravel this mystery from an empirical perspective and show that the knowledge comes from the top-1 predictions of teachers, which also helps us build a potential connection between word- and sequence-level KD. Further, we point out two inherent issues in vanilla word-level KD based on this finding. Firstly, the current objective of KD spreads its focus to whole distributions to learn the knowledge, yet lacks special treatment on the most crucial top-1 information. Secondly, the knowledge is largely covered by the golden information due to the fact that most top-1 predictions of teachers overlap with ground-truth tokens, which further restricts the potential of KD. To address these issues, we propose a novel method named \textbf{T}op-1 \textbf{I}nformation \textbf{E}nhanced \textbf{K}nowledge \textbf{D}istillation (TIE-KD). Specifically, we design a hierarchical ranking loss to enforce the learning of the top-1 information from the teacher. Additionally, we develop an iterative KD procedure to infuse more additional knowledge by distilling on the data without ground-truth targets. Experiments on WMT'14 English-German, WMT'14 English-French and WMT'16 English-Romanian demonstrate that our method can respectively boost Transformer$_{base}$ students by +1.04, +0.60 and +1.11 BLEU scores and significantly outperform the vanilla word-level KD baseline. Besides, our method shows higher generalizability on different teacher-student capacity gaps than existing KD techniques.
Transfer learning has become a popular task adaptation method in the era of foundation models. However, many foundation models require large storage and computing resources, which makes off-the-shelf deployment impractical. Post-training compression techniques such as pruning and quantization can help lower deployment costs. Unfortunately, the resulting performance degradation limits the usability and benefits of such techniques. To close this performance gap, we propose CrAFT, a simple fine-tuning framework that enables effective post-training network compression. In CrAFT, users simply employ the default fine-tuning schedule along with sharpness minimization objective, simultaneously facilitating task adaptation and compression-friendliness. Contrary to the conventional sharpness minimization techniques, which are applied during pretraining, the CrAFT approach adds negligible training overhead as fine-tuning is done in under a couple of minutes or hours with a single GPU. The effectiveness of CrAFT, which is a general-purpose tool that can significantly boost one-shot pruning and post-training quantization, is demonstrated on both convolution-based and attention-based vision foundation models on a variety of target tasks. The code will be made publicly available.
Knowledge distillation conducts an effective model compression method while holding some limitations:(1) the feature based distillation methods only focus on distilling the feature map but are lack of transferring the relation of data examples; (2) the relational distillation methods are either limited to the handcrafted functions for relation extraction, such as L2 norm, or weak in inter- and intra- class relation modeling. Besides, the feature divergence of heterogeneous teacher-student architectures may lead to inaccurate relational knowledge transferring. In this work, we propose a novel training framework named Class-Oriented Relational Self Distillation (CORSD) to address the limitations. The trainable relation networks are designed to extract relation of structured data input, and they enable the whole model to better classify samples by transferring the relational knowledge from the deepest layer of the model to shallow layers. Besides, auxiliary classifiers are proposed to make relation networks capture class-oriented relation that benefits classification task. Experiments demonstrate that CORSD achieves remarkable improvements. Compared to baseline, 3.8%, 1.5% and 4.5% averaged accuracy boost can be observed on CIFAR100, ImageNet and CUB-200-2011, respectively.
Code-trained language models have proven to be highly effective for various code intelligence tasks. However, they can be challenging to train and deploy for many software engineering applications due to computational bottlenecks and memory constraints. Implementing effective strategies to address these issues requires a better understanding of these 'black box' models. In this paper, we perform the first neuron-level analysis for source code models to identify \textit{important} neurons within latent representations. We achieve this by eliminating neurons that are highly similar or irrelevant to the given task. This approach helps us understand which neurons and layers can be eliminated (redundancy analysis) and where important code properties are located within the network (concept analysis). Using redundancy analysis, we make observations relevant to knowledge transfer and model optimization applications. We find that over 95\% of the neurons are redundant with respect to our code intelligence tasks and can be eliminated without significant loss in accuracy. We also discover several subsets of neurons that can make predictions with baseline accuracy. Through concept analysis, we explore the traceability and distribution of human-recognizable concepts within latent code representations which could be used to influence model predictions. We trace individual and subsets of important neurons to specific code properties and identify 'number' neurons, 'string' neurons, and higher-level 'text' neurons for token-level tasks and higher-level concepts important for sentence-level downstream tasks. This also helps us understand how decomposable and transferable task-related features are and can help devise better techniques for transfer learning, model compression, and the decomposition of deep neural networks into modules.
Neural network model compression techniques can address the computation issue of deep neural networks on embedded devices in industrial systems. The guaranteed output error computation problem for neural network compression with quantization is addressed in this paper. A merged neural network is built from a feedforward neural network and its quantized version to produce the exact output difference between two neural networks. Then, optimization-based methods and reachability analysis methods are applied to the merged neural network to compute the guaranteed quantization error. Finally, a numerical example is proposed to validate the applicability and effectiveness of the proposed approach.
Pruning - that is, setting a significant subset of the parameters of a neural network to zero - is one of the most popular methods of model compression. Yet, several recent works have raised the issue that pruning may induce or exacerbate bias in the output of the compressed model. Despite existing evidence for this phenomenon, the relationship between neural network pruning and induced bias is not well-understood. In this work, we systematically investigate and characterize this phenomenon in Convolutional Neural Networks for computer vision. First, we show that it is in fact possible to obtain highly-sparse models, e.g. with less than 10% remaining weights, which do not decrease in accuracy nor substantially increase in bias when compared to dense models. At the same time, we also find that, at higher sparsities, pruned models exhibit higher uncertainty in their outputs, as well as increased correlations, which we directly link to increased bias. We propose easy-to-use criteria which, based only on the uncompressed model, establish whether bias will increase with pruning, and identify the samples most susceptible to biased predictions post-compression.
Previous knowledge distillation methods have shown their impressive performance on model compression tasks, however, it is hard to explain how the knowledge they transferred helps to improve the performance of the student network. In this work, we focus on proposing a knowledge distillation method that has both high interpretability and competitive performance. We first revisit the structure of mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions of input is critical for CNN to perform classification. Furthermore, we demonstrate that this capacity can be obtained and enhanced by transferring class activation maps. Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD). Different from previous KD methods, we explore and present several properties of the knowledge transferred by our method, which not only improve the interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Code is available at: https://github.com/GzyAftermath/CAT-KD.
Many existing studies on knowledge distillation have focused on methods in which a student model mimics a teacher model well. Simply imitating the teacher's knowledge, however, is not sufficient for the student to surpass that of the teacher. We explore a method to harness the knowledge of other students to complement the knowledge of the teacher. We propose deep collective knowledge distillation for model compression, called DCKD, which is a method for training student models with rich information to acquire knowledge from not only their teacher model but also other student models. The knowledge collected from several student models consists of a wealth of information about the correlation between classes. Our DCKD considers how to increase the correlation knowledge of classes during training. Our novel method enables us to create better performing student models for collecting knowledge. This simple yet powerful method achieves state-of-the-art performances in many experiments. For example, for ImageNet, ResNet18 trained with DCKD achieves 72.27\%, which outperforms the pretrained ResNet18 by 2.52\%. For CIFAR-100, the student model of ShuffleNetV1 with DCKD achieves 6.55\% higher top-1 accuracy than the pretrained ShuffleNetV1.