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.
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.
In this paper, we propose an ultrafast automated model compression framework called SeerNet for flexible network deployment. Conventional non-differen-tiable methods discretely search the desirable compression policy based on the accuracy from exhaustively trained lightweight models, and existing differentiable methods optimize an extremely large supernet to obtain the required compressed model for deployment. They both cause heavy computational cost due to the complex compression policy search and evaluation process. On the contrary, we obtain the optimal efficient networks by directly optimizing the compression policy with an accurate performance predictor, where the ultrafast automated model compression for various computational cost constraint is achieved without complex compression policy search and evaluation. Specifically, we first train the performance predictor based on the accuracy from uncertain compression policies actively selected by efficient evolutionary search, so that informative supervision is provided to learn the accurate performance predictor with acceptable cost. Then we leverage the gradient that maximizes the predicted performance under the barrier complexity constraint for ultrafast acquisition of the desirable compression policy, where adaptive update stepsizes with momentum are employed to enhance optimality of the acquired pruning and quantization strategy. Compared with the state-of-the-art automated model compression methods, experimental results on image classification and object detection show that our method achieves competitive accuracy-complexity trade-offs with significant reduction of the search cost.
The recent advances in machine learning and deep neural networks have made them attractive candidates for wireless communications functions such as channel estimation, decoding, and downlink channel state information (CSI) compression. However, most of these neural networks are large and inefficient making it a barrier for deployment in practical wireless systems that require low-latency and low memory footprints for individual network functions. To mitigate these limitations, we propose accelerated and compressed efficient neural networks for massive MIMO CSI feedback. Specifically, we have thoroughly investigated the adoption of network pruning, post-training dynamic range quantization, and weight clustering to optimize CSI feedback compression for massive MIMO systems. Furthermore, we have deployed the proposed model compression techniques on commodity hardware and demonstrated that in order to achieve inference gains, specialized libraries that accelerate computations for sparse neural networks are required. Our findings indicate that there is remarkable value in applying these model compression techniques and the proposed joint pruning and quantization approach reduced model size by 86.5% and inference time by 76.2% with minimal impact to model accuracy. These compression methods are crucial to pave the way for practical adoption and deployments of deep learning-based techniques in commercial wireless systems.
Global channel pruning (GCP) aims to remove a subset of channels (filters) across different layers from a deep model without hurting the performance. Previous works focus on either single task model pruning or simply adapting it to multitask scenario, and still face the following problems when handling multitask pruning: 1) Due to the task mismatch, a well-pruned backbone for classification task focuses on preserving filters that can extract category-sensitive information, causing filters that may be useful for other tasks to be pruned during the backbone pruning stage; 2) For multitask predictions, different filters within or between layers are more closely related and interacted than that for single task prediction, making multitask pruning more difficult. Therefore, aiming at multitask model compression, we propose a Performance-Aware Global Channel Pruning (PAGCP) framework. We first theoretically present the objective for achieving superior GCP, by considering the joint saliency of filters from intra- and inter-layers. Then a sequentially greedy pruning strategy is proposed to optimize the objective, where a performance-aware oracle criterion is developed to evaluate sensitivity of filters to each task and preserve the globally most task-related filters. Experiments on several multitask datasets show that the proposed PAGCP can reduce the FLOPs and parameters by over 60% with minor performance drop, and achieves 1.2x$\sim$3.3x acceleration on both cloud and mobile platforms.
A new golden age in astronomy is upon us, dominated by data. Large astronomical surveys are broadcasting unprecedented rates of information, demanding machine learning as a critical component in modern scientific pipelines to handle the deluge of data. The upcoming Legacy Survey of Space and Time (LSST) of the Vera C. Rubin Observatory will raise the big-data bar for time-domain astronomy, with an expected 10 million alerts per-night, and generating many petabytes of data over the lifetime of the survey. Fast and efficient classification algorithms that can operate in real-time, yet robustly and accurately, are needed for time-critical events where additional resources can be sought for follow-up analyses. In order to handle such data, state-of-the-art deep learning architectures coupled with tools that leverage modern hardware accelerators are essential. We showcase how the use of modern deep compression methods can achieve a $18\times$ reduction in model size, whilst preserving classification performance. We also show that in addition to the deep compression techniques, careful choice of file formats can improve inference latency, and thereby throughput of alerts, on the order of $8\times$ for local processing, and $5\times$ in a live production setting. To test this in a live setting, we deploy this optimised version of the original time-series transformer, t2, into the community alert broking system of FINK on real Zwicky Transient Facility (ZTF) alert data, and compare throughput performance with other science modules that exist in FINK. The results shown herein emphasise the time-series transformer's suitability for real-time classification at LSST scale, and beyond, and introduce deep model compression as a fundamental tool for improving deploy-ability and scalable inference of deep learning models for transient classification.
Knowledge Transfer (KT) achieves competitive performance and is widely used for image classification tasks in model compression and transfer learning. Existing KT works transfer the information from a large model ("teacher") to train a small model ("student") by minimizing the difference of their conditionally independent output distributions. However, these works overlook the high-dimension structural knowledge from the intermediate representations of the teacher, which leads to limited effectiveness, and they are motivated by various heuristic intuitions, which makes it difficult to generalize. This paper proposes a novel Contrastive Knowledge Transfer Framework (CKTF), which enables the transfer of sufficient structural knowledge from the teacher to the student by optimizing multiple contrastive objectives across the intermediate representations between them. Also, CKTF provides a generalized agreement to existing KT techniques and increases their performance significantly by deriving them as specific cases of CKTF. The extensive evaluation shows that CKTF consistently outperforms the existing KT works by 0.04% to 11.59% in model compression and by 0.4% to 4.75% in transfer learning on various models and datasets.
Model compression is a crucial part of deploying neural networks (NNs), especially when the memory and storage of computing devices are limited in many applications. This paper focuses on two model compression techniques: low-rank approximation and weight pruning in neural networks, which are very popular nowadays. However, training NN with low-rank approximation and weight pruning always suffers significant accuracy loss and convergence issues. In this paper, a holistic framework is proposed for model compression from a novel perspective of nonconvex optimization by designing an appropriate objective function. Then, we introduce NN-BCD, a block coordinate descent (BCD) algorithm to solve the nonconvex optimization. One advantage of our algorithm is that an efficient iteration scheme can be derived with closed-form, which is gradient-free. Therefore, our algorithm will not suffer from vanishing/exploding gradient problems. Furthermore, with the Kurdyka-{\L}ojasiewicz (K{\L}) property of our objective function, we show that our algorithm globally converges to a critical point at the rate of O(1/k), where k denotes the number of iterations. Lastly, extensive experiments with tensor train decomposition and weight pruning demonstrate the efficiency and superior performance of the proposed framework. Our code implementation is available at https://github.com/ChenyangLi-97/NN-BCD
The existing model compression methods via structured pruning typically require complicated multi-stage procedures. Each individual stage necessitates numerous engineering efforts and domain-knowledge from the end-users which prevent their wider applications onto broader scenarios. We propose the second generation of Only-Train-Once (OTOv2), which first automatically trains and compresses a general DNN only once from scratch to produce a more compact model with competitive performance without fine-tuning. OTOv2 is automatic and pluggable into various deep learning applications, and requires almost minimal engineering efforts from the users. Methodologically, OTOv2 proposes two major improvements: (i) Autonomy: automatically exploits the dependency of general DNNs, partitions the trainable variables into Zero-Invariant Groups (ZIGs), and constructs the compressed model; and (ii) Dual Half-Space Projected Gradient (DHSPG): a novel optimizer to more reliably solve structured-sparsity problems. Numerically, we demonstrate the generality and autonomy of OTOv2 on a variety of model architectures such as VGG, ResNet, CARN, ConvNeXt, DenseNet and StackedUnets, the majority of which cannot be handled by other methods without extensive handcrafting efforts. Together with benchmark datasets including CIFAR10/100, DIV2K, Fashion-MNIST, SVNH and ImageNet, its effectiveness is validated by performing competitively or even better than the state-of-the-arts. The source code is available at https://github.com/tianyic/only_train_once.
Post-training Neural Network (NN) model compression is an attractive approach for deploying large, memory-consuming models on devices with limited memory resources. In this study, we investigate the rate-distortion tradeoff for NN model compression. First, we suggest a Rotation-Invariant Quantization (RIQ) technique that utilizes a single parameter to quantize the entire NN model, yielding a different rate at each layer, i.e., mixed-precision quantization. Then, we prove that our rotation-invariant approach is optimal in terms of compression. We rigorously evaluate RIQ and demonstrate its capabilities on various models and tasks. For example, RIQ facilitates $\times 19.4$ and $\times 52.9$ compression ratios on pre-trained VGG dense and pruned models, respectively, with $<0.4\%$ accuracy degradation. Code is available in \href{https://github.com/ehaleva/RIQ}{github.com/ehaleva/RIQ}.
Knowledge distillation is an effective method for model compression. However, it is still a challenging topic to apply knowledge distillation to detection tasks. There are two key points resulting in poor distillation performance for detection tasks. One is the serious imbalance between foreground and background features, another one is that small object lacks enough feature representation. To solve the above issues, we propose a new distillation method named dual relation knowledge distillation (DRKD), including pixel-wise relation distillation and instance-wise relation distillation. The pixel-wise relation distillation embeds pixel-wise features in the graph space and applies graph convolution to capture the global pixel relation. By distilling the global pixel relation, the student detector can learn the relation between foreground and background features, and avoid the difficulty of distilling features directly for the feature imbalance issue. Besides, we find that instance-wise relation supplements valuable knowledge beyond independent features for small objects. Thus, the instance-wise relation distillation is designed, which calculates the similarity of different instances to obtain a relation matrix. More importantly, a relation filter module is designed to highlight valuable instance relations. The proposed dual relation knowledge distillation is general and can be easily applied for both one-stage and two-stage detectors. Our method achieves state-of-the-art performance, which improves Faster R-CNN based on ResNet50 from 38.4% to 41.6% mAP and improves RetinaNet based on ResNet50 from 37.4% to 40.3% mAP on COCO 2017.
Real-world data contains a vast amount of multimodal information, among which vision and language are the two most representative modalities. Moreover, increasingly heavier models, \textit{e}.\textit{g}., Transformers, have attracted the attention of researchers to model compression. However, how to compress multimodal models, especially vison-language Transformers, is still under-explored. This paper proposes the \textbf{U}nified and \textbf{P}r\textbf{o}gressive \textbf{P}runing (\textbf{\emph{UPop}}) as a universal vison-language Transformer compression framework, which incorporates 1) unifiedly searching multimodal subnets in a continuous optimization space from the original model, which enables automatic assignment of pruning ratios among compressible modalities and structures; 2) progressively searching and retraining the subnet, which maintains convergence between the search and retrain to attain higher compression ratios. Experiments on various tasks, datasets, and model architectures demonstrate the effectiveness and versatility of the proposed UPop framework. The code is available at https://github.com/sdc17/UPop.
Neural fields, also known as coordinate-based or implicit neural representations, have shown a remarkable capability of representing, generating, and manipulating various forms of signals. For video representations, however, mapping pixel-wise coordinates to RGB colors has shown relatively low compression performance and slow convergence and inference speed. Frame-wise video representation, which maps a temporal coordinate to its entire frame, has recently emerged as an alternative method to represent videos, improving compression rates and encoding speed. While promising, it has still failed to reach the performance of state-of-the-art video compression algorithms. In this work, we propose FFNeRV, a novel method for incorporating flow information into frame-wise representations to exploit the temporal redundancy across the frames in videos inspired by the standard video codecs. Furthermore, we introduce a fully convolutional architecture, enabled by one-dimensional temporal grids, improving the continuity of spatial features. Experimental results show that FFNeRV yields the best performance for video compression and frame interpolation among the methods using frame-wise representations or neural fields. To reduce the model size even further, we devise a more compact convolutional architecture using the group and pointwise convolutions. With model compression techniques, including quantization-aware training and entropy coding, FFNeRV outperforms widely-used standard video codecs (H.264 and HEVC) and performs on par with state-of-the-art video compression algorithms.
Post-training quantization (PTQ), which only requires a tiny dataset for calibration without end-to-end retraining, is a light and practical model compression technique. Recently, several PTQ schemes for vision transformers (ViTs) have been presented; unfortunately, they typically suffer from non-trivial accuracy degradation, especially in low-bit cases. In this paper, we propose RepQ-ViT, a novel PTQ framework for ViTs based on quantization scale reparameterization, to address the above issues. RepQ-ViT decouples the quantization and inference processes, where the former employs complex quantizers and the latter employs scale-reparameterized simplified quantizers. This ensures both accurate quantization and efficient inference, which distinguishes it from existing approaches that sacrifice quantization performance to meet the target hardware. More specifically, we focus on two components with extreme distributions: post-LayerNorm activations with severe inter-channel variation and post-Softmax activations with power-law features, and initially apply channel-wise quantization and log$\sqrt{2}$ quantization, respectively. Then, we reparameterize the scales to hardware-friendly layer-wise quantization and log2 quantization for inference, with only slight accuracy or computational costs. Extensive experiments are conducted on multiple vision tasks with different model variants, proving that RepQ-ViT, without hyperparameters and expensive reconstruction procedures, can outperform existing strong baselines and encouragingly improve the accuracy of 4-bit PTQ of ViTs to a usable level. Code is available at https://github.com/zkkli/RepQ-ViT.
Federated Deep Learning frameworks can be used strategically to monitor Land Use locally and infer environmental impacts globally. Distributed data from across the world would be needed to build a global model for Land Use classification. The need for a Federated approach in this application domain would be to avoid transfer of data from distributed locations and save network bandwidth to reduce communication cost. We use a Federated UNet model for Semantic Segmentation of satellite and street view images. The novelty of the proposed architecture is the integration of Knowledge Distillation to reduce communication cost and response time. The accuracy obtained was above 95% and we also brought in a significant model compression to over 17 times and 62 times for street View and satellite images respectively. Our proposed framework has the potential to be a game-changer in real-time tracking of climate change across the planet.
Approximating radiance fields with volumetric grids is one of promising directions for improving NeRF, represented by methods like Plenoxels and DVGO, which achieve super-fast training convergence and real-time rendering. However, these methods typically require a tremendous storage overhead, costing up to hundreds of megabytes of disk space and runtime memory for a single scene. We address this issue in this paper by introducing a simple yet effective framework, called vector quantized radiance fields (VQRF), for compressing these volume-grid-based radiance fields. We first present a robust and adaptive metric for estimating redundancy in grid models and performing voxel pruning by better exploring intermediate outputs of volumetric rendering. A trainable vector quantization is further proposed to improve the compactness of grid models. In combination with an efficient joint tuning strategy and post-processing, our method can achieve a compression ratio of 100$\times$ by reducing the overall model size to 1 MB with negligible loss on visual quality. Extensive experiments demonstrate that the proposed framework is capable of achieving unrivaled performance and well generalization across multiple methods with distinct volumetric structures, facilitating the wide use of volumetric radiance fields methods in real-world applications. Code Available at \url{https://github.com/AlgoHunt/VQRF}
The problem of broad practical interest in spatiotemporal data analysis, i.e., discovering interpretable dynamic patterns from spatiotemporal data, is studied in this paper. Towards this end, we develop a time-varying reduced-rank vector autoregression (VAR) model whose coefficient matrices are parameterized by low-rank tensor factorization. Benefiting from the tensor factorization structure, the proposed model can simultaneously achieve model compression and pattern discovery. In particular, the proposed model allows one to characterize nonstationarity and time-varying system behaviors underlying spatiotemporal data. To evaluate the proposed model, extensive experiments are conducted on various spatiotemporal data representing different nonlinear dynamical systems, including fluid dynamics, sea surface temperature, USA surface temperature, and NYC taxi trips. Experimental results demonstrate the effectiveness of modeling spatiotemporal data and characterizing spatial/temporal patterns with the proposed model. In the spatial context, the spatial patterns can be automatically extracted and intuitively characterized by the spatial modes. In the temporal context, the complex time-varying system behaviors can be revealed by the temporal modes in the proposed model. Thus, our model lays an insightful foundation for understanding complex spatiotemporal data in real-world dynamical systems. The adapted datasets and Python implementation are publicly available at https://github.com/xinychen/vars.
As a promising solution for model compression, knowledge distillation (KD) has been applied in recommender systems (RS) to reduce inference latency. Traditional solutions first train a full teacher model from the training data, and then transfer its knowledge (\ie \textit{soft labels}) to supervise the learning of a compact student model. However, we find such a standard distillation paradigm would incur serious bias issue -- popular items are more heavily recommended after the distillation. This effect prevents the student model from making accurate and fair recommendations, decreasing the effectiveness of RS. In this work, we identify the origin of the bias in KD -- it roots in the biased soft labels from the teacher, and is further propagated and intensified during the distillation. To rectify this, we propose a new KD method with a stratified distillation strategy. It first partitions items into multiple groups according to their popularity, and then extracts the ranking knowledge within each group to supervise the learning of the student. Our method is simple and teacher-agnostic -- it works on distillation stage without affecting the training of the teacher model. We conduct extensive theoretical and empirical studies to validate the effectiveness of our proposal. We release our code at: https://github.com/chengang95/UnKD.