Network pruning and knowledge distillation are two widely-known model compression methods that efficiently reduce computation cost and model size. A common problem in both pruning and distillation is to determine compressed architecture, i.e., the exact number of filters per layer and layer configuration, in order to preserve most of the original model capacity. In spite of the great advances in existing works, the determination of an excellent architecture still requires human interference or tremendous experimentations. In this paper, we propose an architecture slimming method that automates the layer configuration process. We start from the perspective that the capacity of the over-parameterized model can be largely preserved by finding the minimum number of filters preserving the maximum parameter variance per layer, resulting in a thin architecture. We formulate the determination of compressed architecture as a one-step orthogonal linear transformation, and integrate principle component analysis (PCA), where the variances of filters in the first several projections are maximized. We demonstrate the rationality of our analysis and the effectiveness of the proposed method through extensive experiments. In particular, we show that under the same overall compression rate, the compressed architecture determined by our method shows significant performance gain over baselines after pruning and distillation. Surprisingly, we find that the resulting layer-wise compression rates correspond to the layer sensitivities found by existing works through tremendous experimentations.
Federated edge learning (FEEL) is a promising distributed machine learning (ML) framework to drive edge intelligence applications. However, due to the dynamic wireless environments and the resource limitations of edge devices, communication becomes a major bottleneck. In this work, we propose time-correlated sparsification with hybrid aggregation (TCS-H) for communication-efficient FEEL, which exploits jointly the power of model compression and over-the-air computation. By exploiting the temporal correlations among model parameters, we construct a global sparsification mask, which is identical across devices, and thus enables efficient model aggregation over-the-air. Each device further constructs a local sparse vector to explore its own important parameters, which are aggregated via digital communication with orthogonal multiple access. We further design device scheduling and power allocation algorithms for TCS-H. Experiment results show that, under limited communication resources, TCS-H can achieve significantly higher accuracy compared to the conventional top-K sparsification with orthogonal model aggregation, with both i.i.d. and non-i.i.d. data distributions.
Despite achieving state-of-the-art performance on many NLP tasks, the high energy cost and long inference delay prevent Transformer-based pretrained language models (PLMs) from seeing broader adoption including for edge and mobile computing. Efficient NLP research aims to comprehensively consider computation, time and carbon emission for the entire life-cycle of NLP, including data preparation, model training and inference. In this survey, we focus on the inference stage and review the current state of model compression and acceleration for pretrained language models, including benchmarks, metrics and methodology.
Knowledge distillation is a representative technique for model compression and acceleration, which is important for deploying neural networks on resource limited devices. The knowledge transferred from teacher to student is the mapping of teacher model, or represented by all the input-output pairs. However, in practice the student model only learns from data pairs of the dataset that may be biased, and we think this limits the performance of knowledge distillation. In this paper, we first quantitatively define the uniformity of the sampled data for training, providing a unified view for methods that learn from biased data. Then we evaluate the uniformity on real world dataset and show that existing methods actually improve the uniformity of data. We further introduce two uniformity-oriented methods for rectifying the bias of data for knowledge distillation. Extensive experiments conducted on Face Recognition and Person Re-identification have shown the effectiveness of our method. Moreover, we analyze the sampled data on Face Recognition and show that better balance is achieved between races and between easy and hard samples. And this effect can be also confirmed in training the student model from scratch, resulting in a comparable performance with standard knowledge distillation.
The recent focus on the efficiency of deep neural networks (DNNs) has led to significant work on model compression approaches, of which weight pruning is one of the most popular. At the same time, there is rapidly-growing computational support for efficiently executing the unstructured-sparse models obtained via pruning. Yet, most existing pruning methods minimize just the number of remaining weights, i.e. the size of the model, rather than optimizing for inference time. We address this gap by introducing SPDY, a new compression method which automatically determines layer-wise sparsity targets achieving a desired inference speedup on a given system, while minimizing accuracy loss. SPDY is composed of two new techniques: the first is an efficient dynamic programming algorithm for solving the speedup-constrained layer-wise compression problem assuming a set of given layer-wise sensitivity scores; the second is a local search procedure for determining accurate layer-wise sensitivity scores. Experiments across popular vision and language models show that SPDY guarantees speedups while recovering higher accuracy relative to existing strategies, both for one-shot and gradual pruning scenarios, and is compatible with most existing pruning approaches. We also extend our approach to the recently-proposed task of pruning with very little data, where we achieve the best known accuracy recovery when pruning to the GPU-supported 2:4 sparsity pattern.
Neural networks are more expressive when they have multiple layers. In turn, conventional training methods are only successful if the depth does not lead to numerical issues such as exploding or vanishing gradients, which occur less frequently when the layers are sufficiently wide. However, increasing width to attain greater depth entails the use of heavier computational resources and leads to overparameterized models. These subsequent issues have been partially addressed by model compression methods such as quantization and pruning, some of which relying on normalization-based regularization of the loss function to make the effect of most parameters negligible. In this work, we propose instead to use regularization for preventing neurons from dying or becoming linear, a technique which we denote as jumpstart regularization. In comparison to conventional training, we obtain neural networks that are thinner, deeper, and - most importantly - more parameter-efficient.
In modern neural networks like Transformers, linear layers require significant memory to store activations during backward pass. This study proposes a memory reduction approach to perform backpropagation through linear layers. Since the gradients of linear layers are computed by matrix multiplications, we consider methods for randomized matrix multiplications and demonstrate that they require less memory with a moderate decrease of the test accuracy. Also, we investigate the variance of the gradient estimate induced by the randomized matrix multiplication. We compare this variance with the variance coming from gradient estimation based on the batch of samples. We demonstrate the benefits of the proposed method on the fine-tuning of the pre-trained RoBERTa model on GLUE tasks.
Deep learning models have introduced various intelligent applications to edge devices, such as image classification, speech recognition, and augmented reality. There is an increasing need of training such models on the devices in order to deliver personalized, responsive, and private learning. To address this need, this paper presents a new solution for deploying and training state-of-the-art models on the resource-constrained devices. First, the paper proposes a novel filter-pruning-based model compression method to create lightweight trainable models from large models trained in the cloud, without much loss of accuracy. Second, it proposes a novel knowledge transfer method to enable the on-device model to update incrementally in real time or near real time using incremental learning on new data and enable the on-device model to learn the unseen categories with the help of the in-cloud model in an unsupervised fashion. The results show that 1) our model compression method can remove up to 99.36% parameters of WRN-28-10, while preserving a Top-1 accuracy of over 90% on CIFAR-10; 2) our knowledge transfer method enables the compressed models to achieve more than 90% accuracy on CIFAR-10 and retain good accuracy on old categories; 3) it allows the compressed models to converge within real time (three to six minutes) on the edge for incremental learning tasks; 4) it enables the model to classify unseen categories of data (78.92% Top-1 accuracy) that it is never trained with.
Model compression methods can reduce model complexity on the premise of maintaining acceptable performance, and thus promote the application of deep neural networks under resource constrained environments. Despite their great success, the selection of suitable compression methods and design of details of the compression scheme are difficult, requiring lots of domain knowledge as support, which is not friendly to non-expert users. To make more users easily access to the model compression scheme that best meet their needs, in this paper, we propose AutoMC, an effective automatic tool for model compression. AutoMC builds the domain knowledge on model compression to deeply understand the characteristics and advantages of each compression method under different settings. In addition, it presents a progressive search strategy to efficiently explore pareto optimal compression scheme according to the learned prior knowledge combined with the historical evaluation information. Extensive experimental results show that AutoMC can provide satisfying compression schemes within short time, demonstrating the effectiveness of AutoMC.
Model compression techniques are receiving increasing attention; however, the effect of compression on model fairness is still under explored. This is the first paper to examine the effect of distillation and pruning on the toxicity and bias of generative language models. We test Knowledge Distillation and Pruning methods on the GPT2 model and found a consistent pattern of toxicity and bias reduction after model distillation; this result can be potentially interpreted by existing line of research which describes model compression as a regularization technique; our work not only serves as a reference for safe deployment of compressed models, but also extends the discussion of "compression as regularization" into the setting of neural LMs, and hints at the possibility of using compression to develop fairer models.
Recently, large pre-trained models have significantly improved the performance of various Natural LanguageProcessing (NLP) tasks but they are expensive to serve due to long serving latency and large memory usage. To compress these models, knowledge distillation has attracted an increasing amount of interest as one of the most effective methods for model compression. However, existing distillation methods have not yet addressed the unique challenges of model serving in datacenters, such as handling fast evolving models, considering serving performance, and optimizing for multiple objectives. To solve these problems, we propose AutoDistill, an end-to-end model distillation framework integrating model architecture exploration and multi-objective optimization for building hardware-efficient NLP pre-trained models. We use Bayesian Optimization to conduct multi-objective Neural Architecture Search for selecting student model architectures. The proposed search comprehensively considers both prediction accuracy and serving latency on target hardware. The experiments on TPUv4i show the finding of seven model architectures with better pre-trained accuracy (up to 3.2% higher) and lower inference latency (up to 1.44x faster) than MobileBERT. By running downstream NLP tasks in the GLUE benchmark, the model distilled for pre-training by AutoDistill with 28.5M parameters achieves an 81.69 average score, which is higher than BERT_BASE, DistillBERT, TinyBERT, NAS-BERT, and MobileBERT. The most compact model found by AutoDistill contains only 20.6M parameters but still outperform BERT_BASE(109M), DistillBERT(67M), TinyBERT(67M), and MobileBERT(25.3M) regarding the average GLUE score. By evaluating on SQuAD, a model found by AutoDistill achieves an 88.4% F1 score with 22.8M parameters, which reduces parameters by more than 62% while maintaining higher accuracy than DistillBERT, TinyBERT, and NAS-BERT.
In recent years, neural signed distance function (SDF) has become one of the most effective representation methods for 3D models. By learning continuous SDFs in 3D space, neural networks can predict the distance from a given query space point to its closest object surface,whose positive and negative signs denote inside and outside of the object, respectively. Training a specific network for each 3D model, which individually embeds its shape, can realize compressed representation of objects by storing fewer network (and possibly latent) parameters. Consequently, reconstruction through network inference and surface recovery can be achieved. In this paper, we propose an SDF prediction network using explicit key spheres as input. Key spheres are extracted from the internal space of objects, whose centers either have relatively larger SDF values (sphere radii), or are located at essential positions. By inputting the spatial information of multiple spheres which imply different local shapes, the proposed method can significantly improve the reconstruction accuracy with a negligible storage cost. Compared to previous works, our method achieves the high-fidelity and high-compression 3D object coding and reconstruction. Experiments conducted on three datasets verify the superior performance of our method.
Deploying TinyML models on low-cost IoT hardware is very challenging, due to limited device memory capacity. Neural processing unit (NPU) hardware address the memory challenge by using model compression to exploit weight quantization and sparsity to fit more parameters in the same footprint. However, designing compressible neural networks (NNs) is challenging, as it expands the design space across which we must make balanced trade-offs. This paper demonstrates Unified DNAS for Compressible (UDC) NNs, which explores a large search space to generate state-of-the-art compressible NNs for NPU. ImageNet results show UDC networks are up to $3.35\times$ smaller (iso-accuracy) or 6.25% more accurate (iso-model size) than previous work.
As the training of giant dense models hits the boundary on the availability and capability of the hardware resources today, Mixture-of-Experts (MoE) models become one of the most promising model architectures due to their significant training cost reduction compared to a quality-equivalent dense model. Its training cost saving is demonstrated from encoder-decoder models (prior works) to a 5x saving for auto-aggressive language models (this work along with parallel explorations). However, due to the much larger model size and unique architecture, how to provide fast MoE model inference remains challenging and unsolved, limiting its practical usage. To tackle this, we present DeepSpeed-MoE, an end-to-end MoE training and inference solution as part of the DeepSpeed library, including novel MoE architecture designs and model compression techniques that reduce MoE model size by up to 3.7x, and a highly optimized inference system that provides 7.3x better latency and cost compared to existing MoE inference solutions. DeepSpeed-MoE offers an unprecedented scale and efficiency to serve massive MoE models with up to 4.5x faster and 9x cheaper inference compared to quality-equivalent dense models. We hope our innovations and systems help open a promising path to new directions in the large model landscape, a shift from dense to sparse MoE models, where training and deploying higher-quality models with fewer resources becomes more widely possible.
Speech recognition on smart devices is challenging owing to the small memory footprint. Hence small size ASR models are desirable. With the use of popular transducer-based models, it has become possible to practically deploy streaming speech recognition models on small devices [1]. Recently, the two-pass model [2] combining RNN-T and LAS modules has shown exceptional performance for streaming on-device speech recognition. In this work, we propose a simple and effective approach to reduce the size of the two-pass model for memory-constrained devices. We employ a popular knowledge distillation approach in three stages using the Teacher-Student training technique. In the first stage, we use a trained RNN-T model as a teacher model and perform knowledge distillation to train the student RNN-T model. The second stage uses the shared encoder and trains a LAS rescorer for student model using the trained RNN-T+LAS teacher model. Finally, we perform deep-finetuning for the student model with a shared RNN-T encoder, RNN-T decoder, and LAS rescorer. Our experimental results on standard LibriSpeech dataset show that our system can achieve a high compression rate of 55% without significant degradation in the WER compared to the two-pass teacher model.
With the continuous development of neural networks for computer vision tasks, more and more network architectures have achieved outstanding success. As one of the most advanced neural network architectures, DenseNet shortcuts all feature maps to solve the model depth problem. Although this network architecture has excellent accuracy with low parameters, it requires an excessive inference time. To solve this problem, HarDNet reduces the connections between the feature maps, making the remaining connections resemble harmonic waves. However, this compression method may result in a decrease in the model accuracy and an increase in the parameters and model size. This network architecture may reduce the memory access time, but its overall performance can still be improved. Therefore, we propose a new network architecture, ThreshNet, using a threshold mechanism to further optimize the connection method. Different numbers of connections for different convolution layers are discarded to accelerate the inference of the network. The proposed network has been evaluated with image classification using CIFAR 10 and SVHN datasets under platforms of NVIDIA RTX 3050 and Raspberry Pi 4. The experimental results show that, compared with HarDNet68, GhostNet, MobileNetV2, ShuffleNet, and EfficientNet, the inference time of the proposed ThreshNet79 is 5%, 9%, 10%, 18%, and 20% faster, respectively. The number of parameters of ThreshNet95 is 55% less than that of HarDNet85. The new model compression and model acceleration methods can speed up the inference time, enabling network models to operate on mobile devices.
Deep neural networks have proved hugely successful, achieving human-like performance on a variety of tasks. However, they are also computationally expensive, which has motivated the development of model compression techniques which reduce the resource consumption associated with deep learning models. Nevertheless, recent studies have suggested that model compression can have an adverse effect on algorithmic fairness, amplifying existing biases in machine learning models. With this project we aim to extend those studies to the context of facial expression recognition. To do that, we set up a neural network classifier to perform facial expression recognition and implement several model compression techniques on top of it. We then run experiments on two facial expression datasets, namely the Extended Cohn-Kanade Dataset (CK+DB) and the Real-World Affective Faces Database (RAF-DB), to examine the individual and combined effect that compression techniques have on the model size, accuracy and fairness. Our experimental results show that: (i) Compression and quantisation achieve significant reduction in model size with minimal impact on overall accuracy for both CK+DB and RAF-DB; (ii) in terms of model accuracy, the classifier trained and tested on RAF-DB seems more robust to compression compared to the CK+ DB; (iii) for RAF-DB, the different compression strategies do not seem to increase the gap in predictive performance across the sensitive attributes of gender, race and age which is in contrast with the results on the CK+DB, where compression seems to amplify existing biases for gender. We analyse the results and discuss the potential reasons for our findings.
Knowledge distillation has made remarkable achievements in model compression. However, most existing methods require the original training data, which is usually unavailable due to privacy and security issues. In this paper, we propose a conditional generative data-free knowledge distillation (CGDD) framework for training lightweight networks without any training data. This method realizes efficient knowledge distillation based on conditional image generation. Specifically, we treat the preset labels as ground truth to train a conditional generator in a semi-supervised manner. The trained generator can produce specified classes of training images. For training the student network, we force it to extract the knowledge hidden in teacher feature maps, which provide crucial cues for the learning process. Moreover, an adversarial training framework for promoting distillation performance is constructed by designing several loss functions. This framework helps the student model to explore larger data space. To demonstrate the effectiveness of the proposed method, we conduct extensive experiments on different datasets. Compared with other data-free works, our work obtains state-of-the-art results on CIFAR100, Caltech101, and different versions of ImageNet datasets. The codes will be released.
In the last decade, many deep learning models have been well trained and made a great success in various fields of machine intelligence, especially for computer vision and natural language processing. To better leverage the potential of these well-trained models in intra-domain or cross-domain transfer learning situations, knowledge distillation (KD) and domain adaptation (DA) are proposed and become research highlights. They both aim to transfer useful information from a well-trained model with original training data. However, the original data is not always available in many cases due to privacy, copyright or confidentiality. Recently, the data-free knowledge transfer paradigm has attracted appealing attention as it deals with distilling valuable knowledge from well-trained models without requiring to access to the training data. In particular, it mainly consists of the data-free knowledge distillation (DFKD) and source data-free domain adaptation (SFDA). On the one hand, DFKD aims to transfer the intra-domain knowledge of original data from a cumbersome teacher network to a compact student network for model compression and efficient inference. On the other hand, the goal of SFDA is to reuse the cross-domain knowledge stored in a well-trained source model and adapt it to a target domain. In this paper, we provide a comprehensive survey on data-free knowledge transfer from the perspectives of knowledge distillation and unsupervised domain adaptation, to help readers have a better understanding of the current research status and ideas. Applications and challenges of the two areas are briefly reviewed, respectively. Furthermore, we provide some insights to the subject of future research.
Emerging edge applications require both a fast response latency and complex processing. This is infeasible without expensive hardware that can process complex operations -- such as object detection -- within a short time. Many approach this problem by addressing the complexity of the models -- via model compression, pruning and quantization -- or compressing the input. In this paper, we propose a different perspective when addressing the performance challenges. Croesus is a multi-stage approach to edge-cloud systems that provides the ability to find the balance between accuracy and performance. Croesus consists of two stages (that can be generalized to multiple stages): an initial and a final stage. The initial stage performs the computation in real-time using approximate/best-effort computation at the edge. The final stage performs the full computation at the cloud, and uses the results to correct any errors made at the initial stage. In this paper, we demonstrate the implications of such an approach on a video analytics use-case and show how multi-stage processing yields a better balance between accuracy and performance. Moreover, we study the safety of multi-stage transactions via two proposals: multi-stage serializability (MS-SR) and multi-stage invariant confluence with Apologies (MS-IA).