Probabilistic circuits (PCs) are a tractable representation of probability distributions allowing for exact and efficient computation of likelihoods and marginals. There has been significant recent progress on improving the scale and expressiveness of PCs. However, PC training performance plateaus as model size increases. We discover that most capacity in existing large PC structures is wasted: fully-connected parameter layers are only sparsely used. We propose two operations: pruning and growing, that exploit the sparsity of PC structures. Specifically, the pruning operation removes unimportant sub-networks of the PC for model compression and comes with theoretical guarantees. The growing operation increases model capacity by increasing the size of the latent space. By alternatingly applying pruning and growing, we increase the capacity that is meaningfully used, allowing us to significantly scale up PC learning. Empirically, our learner achieves state-of-the-art likelihoods on MNIST-family image datasets and on Penn Tree Bank language data compared to other PC learners and less tractable deep generative models such as flow-based models and variational autoencoders (VAEs).
Knowledge distillation (KD) has been a ubiquitous method for model compression to strengthen the capability of a lightweight model with the transferred knowledge from the teacher. In particular, KD has been employed in quantization-aware training (QAT) of Transformer encoders like BERT to improve the accuracy of the student model with the reduced-precision weight parameters. However, little is understood about which of the various KD approaches best fits the QAT of Transformers. In this work, we provide an in-depth analysis of the mechanism of KD on attention recovery of quantized large Transformers. In particular, we reveal that the previously adopted MSE loss on the attention score is insufficient for recovering the self-attention information. Therefore, we propose two KD methods; attention-map and attention-output losses. Furthermore, we explore the unification of both losses to address task-dependent preference between attention-map and output losses. The experimental results on various Transformer encoder models demonstrate that the proposed KD methods achieve state-of-the-art accuracy for QAT with sub-2-bit weight quantization.
Transformer-based self-supervised models have achieved remarkable success in speech processing, but their large size and high inference cost present significant challenges for real-world deployment. While numerous compression techniques have been proposed, inconsistent evaluation metrics make it difficult to compare their practical effectiveness. In this work, we conduct a comprehensive study of four common compression methods, including weight pruning, head pruning, low-rank approximation, and knowledge distillation on self-supervised speech Transformers. We evaluate each method under three key metrics: parameter count, multiply-accumulate operations, and real-time factor. Results show that each method offers distinct advantages. In addition, we contextualize recent compression techniques, comparing DistilHuBERT, FitHuBERT, LightHuBERT, ARMHuBERT, and STaRHuBERT under the same framework, offering practical guidance on compression for deployment.
Transformer-based language models have become the standard approach to solving natural language processing tasks. However, industry adoption usually requires the maximum throughput to comply with certain latency constraints that prevents Transformer models from being used in production. To address this gap, model compression techniques such as quantization and pruning may be used to improve inference efficiency. However, these compression techniques require specialized software to apply and deploy at scale. In this work, we propose a new pipeline for creating and running Fast Transformer models on CPUs, utilizing hardware-aware pruning, knowledge distillation, quantization, and our own Transformer inference runtime engine with optimized kernels for sparse and quantized operators. We demonstrate the efficiency of our pipeline by creating a Fast DistilBERT model showing minimal accuracy loss on the question-answering SQuADv1.1 benchmark, and throughput results under typical production constraints and environments. Our results outperform existing state-of-the-art Neural Magic's DeepSparse runtime performance by up to 50% and up to 4.1x performance speedup over ONNX Runtime. Source code is publicly available at https://github.com/intel/intel-extension-for-transformers.
Mixup is a popular data augmentation technique based on creating new samples by linear interpolation between two given data samples, to improve both the generalization and robustness of the trained model. Knowledge distillation (KD), on the other hand, is widely used for model compression and transfer learning, which involves using a larger network's implicit knowledge to guide the learning of a smaller network. At first glance, these two techniques seem very different, however, we found that "smoothness" is the connecting link between the two and is also a crucial attribute in understanding KD's interplay with mixup. Although many mixup variants and distillation methods have been proposed, much remains to be understood regarding the role of a mixup in knowledge distillation. In this paper, we present a detailed empirical study on various important dimensions of compatibility between mixup and knowledge distillation. We also scrutinize the behavior of the networks trained with a mixup in the light of knowledge distillation through extensive analysis, visualizations, and comprehensive experiments on image classification. Finally, based on our findings, we suggest improved strategies to guide the student network to enhance its effectiveness. Additionally, the findings of this study provide insightful suggestions to researchers and practitioners that commonly use techniques from KD. Our code is available at https://github.com/hchoi71/MIX-KD.
Generative Pre-trained Transformer models, known as GPT or OPT, set themselves apart through breakthrough performance across complex language modelling tasks, but also by their extremely high computational and storage costs. Specifically, due to their massive size, even inference for large, highly-accurate GPT models may require multiple performant GPUs, which limits the usability of such models. While there is emerging work on relieving this pressure via model compression, the applicability and performance of existing compression techniques is limited by the scale and complexity of GPT models. In this paper, we address this challenge, and propose GPTQ, a new one-shot weight quantization method based on approximate second-order information, that is both highly-accurate and highly-efficient. Specifically, GPTQ can quantize GPT models with 175 billion parameters in approximately four GPU hours, reducing the bitwidth down to 3 or 4 bits per weight, with negligible accuracy degradation relative to the uncompressed baseline. Our method more than doubles the compression gains relative to previously-proposed one-shot quantization methods, preserving accuracy, allowing us for the first time to execute an 175 billion-parameter model inside a single GPU for generative inference. Moreover, we also show that our method can still provide reasonable accuracy in the extreme quantization regime, in which weights are quantized to 2-bit or even ternary quantization levels. We show experimentally that these improvements can be leveraged for end-to-end inference speedups over FP16, of around 3.25x when using high-end GPUs (NVIDIA A100) and 4.5x when using more cost-effective ones (NVIDIA A6000). The implementation is available at https://github.com/IST-DASLab/gptq.
Transformer-based pre-trained language models (PLMs) mostly suffer from excessive overhead despite their advanced capacity. For resource-constrained devices, there is an urgent need for a spatially and temporally efficient model which retains the major capacity of PLMs. However, existing statically compressed models are unaware of the diverse complexities between input instances, potentially resulting in redundancy and inadequacy for simple and complex inputs. Also, miniature models with early exiting encounter challenges in the trade-off between making predictions and serving the deeper layers. Motivated by such considerations, we propose a collaborative optimization for PLMs that integrates static model compression and dynamic inference acceleration. Specifically, the PLM is slenderized in width while the depth remains intact, complementing layer-wise early exiting to speed up inference dynamically. To address the trade-off of early exiting, we propose a joint training approach that calibrates slenderization and preserves contributive structures to each exit instead of only the final layer. Experiments are conducted on GLUE benchmark and the results verify the Pareto optimality of our approach at high compression and acceleration rate with 1/8 parameters and 1/19 FLOPs of BERT.
A deeper network structure generally handles more complicated non-linearity and performs more competitively. Nowadays, advanced network designs often contain a large number of repetitive structures (e.g., Transformer). They empower the network capacity to a new level but also increase the model size inevitably, which is unfriendly to either model restoring or transferring. In this study, we are the first to investigate the representative potential of fixed random weights with limited unique values by learning diverse masks and introduce the Parameter-Efficient Masking Networks (PEMN). It also naturally leads to a new paradigm for model compression to diminish the model size. Concretely, motivated by the repetitive structures in modern neural networks, we utilize one random initialized layer, accompanied with different masks, to convey different feature mappings and represent repetitive network modules. Therefore, the model can be expressed as \textit{one-layer} with a bunch of masks, which significantly reduce the model storage cost. Furthermore, we enhance our strategy by learning masks for a model filled by padding a given random weights vector. In this way, our method can further lower the space complexity, especially for models without many repetitive architectures. We validate the potential of PEMN learning masks on random weights with limited unique values and test its effectiveness for a new compression paradigm based on different network architectures. Code is available at https://github.com/yueb17/PEMN
Layer-wise distillation is a powerful tool to compress large models (i.e. teacher models) into small ones (i.e., student models). The student distills knowledge from the teacher by mimicking the hidden representations of the teacher at every intermediate layer. However, layer-wise distillation is difficult. Since the student has a smaller model capacity than the teacher, it is often under-fitted. Furthermore, the hidden representations of the teacher contain redundant information that the student does not necessarily need for the target task's learning. To address these challenges, we propose a novel Task-aware layEr-wise Distillation (TED). TED designs task-aware filters to align the hidden representations of the student and the teacher at each layer. The filters select the knowledge that is useful for the target task from the hidden representations. As such, TED reduces the knowledge gap between the two models and helps the student to fit better on the target task. We evaluate TED in two scenarios: continual pre-training and fine-tuning. TED demonstrates significant and consistent improvements over existing distillation methods in both scenarios. Code is available at https://github.com/cliang1453/task-aware-distillation.
Lighter and faster image restoration (IR) models are crucial for the deployment on resource-limited devices. Binary neural network (BNN), one of the most promising model compression methods, can dramatically reduce the computations and parameters of full-precision convolutional neural networks (CNN). However, there are different properties between BNN and full-precision CNN, and we can hardly use the experience of designing CNN to develop BNN. In this study, we reconsider components in binary convolution, such as residual connection, BatchNorm, activation function, and structure, for IR tasks. We conduct systematic analyses to explain each component's role in binary convolution and discuss the pitfalls. Specifically, we find that residual connection can reduce the information loss caused by binarization; BatchNorm can solve the value range gap between residual connection and binary convolution; The position of the activation function dramatically affects the performance of BNN. Based on our findings and analyses, we design a simple yet efficient basic binary convolution unit (BBCU). Furthermore, we divide IR networks into four parts and specially design variants of BBCU for each part to explore the benefit of binarizing these parts. We conduct experiments on different IR tasks, and our BBCU significantly outperforms other BNNs and lightweight models, which shows that BBCU can serve as a basic unit for binarized IR networks. All codes and models will be released.
Vision Transformers are increasingly embedded in industrial systems due to their superior performance, but their memory and power requirements make deploying them to edge devices a challenging task. Hence, model compression techniques are now widely used to deploy models on edge devices as they decrease the resource requirements and make model inference very fast and efficient. But their reliability and robustness from a security perspective is another major issue in safety-critical applications. Adversarial attacks are like optical illusions for ML algorithms and they can severely impact the accuracy and reliability of models. In this work we investigate the transferability of adversarial samples across the SOTA Vision Transformer models across 3 SOTA compressed versions and infer the effects different compression techniques have on adversarial attacks.
On-device session-based recommendation systems have been achieving increasing attention on account of the low energy/resource consumption and privacy protection while providing promising recommendation performance. To fit the powerful neural session-based recommendation models in resource-constrained mobile devices, tensor-train decomposition and its variants have been widely applied to reduce memory footprint by decomposing the embedding table into smaller tensors, showing great potential in compressing recommendation models. However, these model compression techniques significantly increase the local inference time due to the complex process of generating index lists and a series of tensor multiplications to form item embeddings, and the resultant on-device recommender fails to provide real-time response and recommendation. To improve the online recommendation efficiency, we propose to learn compositional encoding-based compact item representations. Specifically, each item is represented by a compositional code that consists of several codewords, and we learn embedding vectors to represent each codeword instead of each item. Then the composition of the codeword embedding vectors from different embedding matrices (i.e., codebooks) forms the item embedding. Since the size of codebooks can be extremely small, the recommender model is thus able to fit in resource-constrained devices and meanwhile can save the codebooks for fast local inference.Besides, to prevent the loss of model capacity caused by compression, we propose a bidirectional self-supervised knowledge distillation framework. Extensive experimental results on two benchmark datasets demonstrate that compared with existing methods, the proposed on-device recommender not only achieves an 8x inference speedup with a large compression ratio but also shows superior recommendation performance.
We present a systematic study of domain generalization (DG) for tiny neural networks. This problem is critical to on-device machine learning applications but has been overlooked in the literature where research has been merely focused on large models. Tiny neural networks have much fewer parameters and lower complexity and therefore should not be trained the same way as their large counterparts for DG applications. By conducting extensive experiments, we find that knowledge distillation (KD), a well-known technique for model compression, is much better for tackling the on-device DG problem than conventional DG methods. Another interesting observation is that the teacher-student gap on out-of-distribution data is bigger than that on in-distribution data, which highlights the capacity mismatch issue as well as the shortcoming of KD. We further propose a method called out-of-distribution knowledge distillation (OKD) where the idea is to teach the student how the teacher handles out-of-distribution data synthesized via disruptive data augmentation. Without adding any extra parameter to the model -- hence keeping the deployment cost unchanged -- OKD significantly improves DG performance for tiny neural networks in a variety of on-device DG scenarios for image and speech applications. We also contribute a scalable approach for synthesizing visual domain shifts, along with a new suite of DG datasets to complement existing testbeds.
Data-free quantization can potentially address data privacy and security concerns in model compression, and thus has been widely investigated. Recently, PSAQ-ViT designs a relative value metric, patch similarity, to generate data from pre-trained vision transformers (ViTs), achieving the first attempt at data-free quantization for ViTs. In this paper, we propose PSAQ-ViT V2, a more accurate and general data-free quantization framework for ViTs, built on top of PSAQ-ViT. More specifically, following the patch similarity metric in PSAQ-ViT, we introduce an adaptive teacher-student strategy, which facilitates the constant cyclic evolution of the generated samples and the quantized model (student) in a competitive and interactive fashion under the supervision of the full-precision model (teacher), thus significantly improving the accuracy of the quantized model. Moreover, without the auxiliary category guidance, we employ the task- and model-independent prior information, making the general-purpose scheme compatible with a broad range of vision tasks and models. Extensive experiments are conducted on various models on image classification, object detection, and semantic segmentation tasks, and PSAQ-ViT V2, with the naive quantization strategy and without access to real-world data, consistently achieves competitive results, showing potential as a powerful baseline on data-free quantization for ViTs. For instance, with Swin-S as the (backbone) model, 8-bit quantization reaches 82.13 top-1 accuracy on ImageNet, 50.9 box AP and 44.1 mask AP on COCO, and 47.2 mIoU on ADE20K. We hope that accurate and general PSAQ-ViT V2 can serve as a potential and practice solution in real-world applications involving sensitive data. Code is released and merged at: https://github.com/zkkli/PSAQ-ViT.
As real-world graphs expand in size, larger GNN models with billions of parameters are deployed. High parameter count in such models makes training and inference on graphs expensive and challenging. To reduce the computational and memory costs of GNNs, optimization methods such as pruning the redundant nodes and edges in input graphs have been commonly adopted. However, model compression, which directly targets the sparsification of model layers, has been mostly limited to traditional Deep Neural Networks (DNNs) used for tasks such as image classification and object detection. In this paper, we utilize two state-of-the-art model compression methods (1) train and prune and (2) sparse training for the sparsification of weight layers in GNNs. We evaluate and compare the efficiency of both methods in terms of accuracy, training sparsity, and training FLOPs on real-world graphs. Our experimental results show that on the ia-email, wiki-talk, and stackoverflow datasets for link prediction, sparse training with much lower training FLOPs achieves a comparable accuracy with the train and prune method. On the brain dataset for node classification, sparse training uses a lower number FLOPs (less than 1/7 FLOPs of train and prune method) and preserves a much better accuracy performance under extreme model sparsity.
Device Model Generalization (DMG) is a practical yet under-investigated research topic for on-device machine learning applications. It aims to improve the generalization ability of pre-trained models when deployed on resource-constrained devices, such as improving the performance of pre-trained cloud models on smart mobiles. While quite a lot of works have investigated the data distribution shift across clouds and devices, most of them focus on model fine-tuning on personalized data for individual devices to facilitate DMG. Despite their promising, these approaches require on-device re-training, which is practically infeasible due to the overfitting problem and high time delay when performing gradient calculation on real-time data. In this paper, we argue that the computational cost brought by fine-tuning can be rather unnecessary. We consequently present a novel perspective to improving DMG without increasing computational cost, i.e., device-specific parameter generation which directly maps data distribution to parameters. Specifically, we propose an efficient Device-cloUd collaborative parametErs generaTion framework DUET. DUET is deployed on a powerful cloud server that only requires the low cost of forwarding propagation and low time delay of data transmission between the device and the cloud. By doing so, DUET can rehearse the device-specific model weight realizations conditioned on the personalized real-time data for an individual device. Importantly, our DUET elegantly connects the cloud and device as a 'duet' collaboration, frees the DMG from fine-tuning, and enables a faster and more accurate DMG paradigm. We conduct an extensive experimental study of DUET on three public datasets, and the experimental results confirm our framework's effectiveness and generalisability for different DMG tasks.
Edge devices can benefit remarkably from federated learning due to their distributed nature; however, their limited resource and computing power poses limitations in deployment. A possible solution to this problem is to utilize off-the-shelf sparse learning algorithms at the clients to meet their resource budget. However, such naive deployment in the clients causes significant accuracy degradation, especially for highly resource-constrained clients. In particular, our investigations reveal that the lack of consensus in the sparsity masks among the clients may potentially slow down the convergence of the global model and cause a substantial accuracy drop. With these observations, we present \textit{federated lottery aware sparsity hunting} (FLASH), a unified sparse learning framework for training a sparse sub-model that maintains the performance under ultra-low parameter density while yielding proportional communication benefits. Moreover, given that different clients may have different resource budgets, we present \textit{hetero-FLASH} where clients can take different density budgets based on their device resource limitations instead of supporting only one target parameter density. Experimental analysis on diverse models and datasets shows the superiority of FLASH in closing the gap with an unpruned baseline while yielding up to $\mathord{\sim}10.1\%$ improved accuracy with $\mathord{\sim}10.26\times$ fewer communication, compared to existing alternatives, at similar hyperparameter settings. Code is available at \url{https://github.com/SaraBabakN/flash_fl}.
We consider the problem of model compression for deep neural networks (DNNs) in the challenging one-shot/post-training setting, in which we are given an accurate trained model, and must compress it without any retraining, based only on a small amount of calibration input data. This problem has become popular in view of the emerging software and hardware support for executing models compressed via pruning and/or quantization with speedup, and well-performing solutions have been proposed independently for both compression approaches. In this paper, we introduce a new compression framework which covers both weight pruning and quantization in a unified setting, is time- and space-efficient, and considerably improves upon the practical performance of existing post-training methods. At the technical level, our approach is based on an exact and efficient realization of the classical Optimal Brain Surgeon (OBS) framework of [LeCun, Denker, and Solla, 1990] extended to also cover weight quantization at the scale of modern DNNs. From the practical perspective, our experimental results show that it can improve significantly upon the compression-accuracy trade-offs of existing post-training methods, and that it can enable the accurate compound application of both pruning and quantization in a post-training setting.
The design of an effective multi-bit watermarking algorithm hinges upon finding a good trade-off between the three fundamental requirements forming the watermarking trade-off triangle, namely, robustness against network modifications, payload, and unobtrusiveness, ensuring minimal impact on the performance of the watermarked network. In this paper, we first revisit the nature of the watermarking trade-off triangle for the DNN case, then we exploit our findings to propose a white-box, multi-bit watermarking method achieving very large payload and strong robustness against network modification. In the proposed system, the weights hosting the watermark are set prior to training, making sure that their amplitude is large enough to bear the target payload and survive network modifications, notably retraining, and are left unchanged throughout the training process. The distribution of the weights carrying the watermark is theoretically optimised to ensure the secrecy of the watermark and make sure that the watermarked weights are indistinguishable from the non-watermarked ones. The proposed method can achieve outstanding performance, with no significant impact on network accuracy, including robustness against network modifications, retraining and transfer learning, while ensuring a payload which is out of reach of state of the art methods achieving a lower - or at most comparable - robustness.
The size of deep learning models in artificial intelligence (AI) software is increasing rapidly, which hinders the large-scale deployment on resource-restricted devices (e.g., smartphones). To mitigate this issue, AI software compression plays a crucial role, which aims to compress model size while keeping high performance. However, the intrinsic defects in the big model may be inherited by the compressed one. Such defects may be easily leveraged by attackers, since the compressed models are usually deployed in a large number of devices without adequate protection. In this paper, we try to address the safe model compression problem from a safety-performance co-optimization perspective. Specifically, inspired by the test-driven development (TDD) paradigm in software engineering, we propose a test-driven sparse training framework called SafeCompress. By simulating the attack mechanism as the safety test, SafeCompress can automatically compress a big model to a small one following the dynamic sparse training paradigm. Further, considering a representative attack, i.e., membership inference attack (MIA), we develop a concrete safe model compression mechanism, called MIA-SafeCompress. Extensive experiments are conducted to evaluate MIA-SafeCompress on five datasets for both computer vision and natural language processing tasks. The results verify the effectiveness and generalization of our method. We also discuss how to adapt SafeCompress to other attacks besides MIA, demonstrating the flexibility of SafeCompress.