Knowledge distillation extracts general knowledge from a pre-trained teacher network and provides guidance to a target student network. Most studies manually tie intermediate features of the teacher and student, and transfer knowledge through pre-defined links. However, manual selection often constructs ineffective links that limit the improvement from the distillation. There has been an attempt to address the problem, but it is still challenging to identify effective links under practical scenarios. In this paper, we introduce an effective and efficient feature distillation method utilizing all the feature levels of the teacher without manually selecting the links. Specifically, our method utilizes an attention-based meta-network that learns relative similarities between features, and applies identified similarities to control distillation intensities of all possible pairs. As a result, our method determines competent links more efficiently than the previous approach and provides better performance on model compression and transfer learning tasks. Further qualitative analyses and ablative studies describe how our method contributes to better distillation. The implementation code is available at github.com/clovaai/attention-feature-distillation.
Deep learning approaches have achieved unprecedented performance in visual recognition tasks such as object detection and pose estimation. However, state-of-the-art models have millions of parameters represented as floats which make them computationally expensive and constrain their deployment on hardware such as mobile phones and IoT nodes. Most commonly, activations of deep neural networks tend to be sparse thus proving that models are over parametrized with redundant neurons. Model compression techniques, such as pruning and quantization, have recently shown promising results by improving model complexity with little loss in performance. In this work, we extended pruning, a compression technique that discards unnecessary model connections, and weight sharing techniques for the task of object detection. With our approach, we are able to compress a state-of-the-art object detection model by 30.0% without a loss in performance. We also show that our compressed model can be easily initialized with existing pre-trained weights, and thus is able to fully utilize published state-of-the-art model zoos.
Model compression is an essential technique for deploying deep neural networks (DNNs) on power and memory-constrained resources. However, existing model-compression methods often rely on human expertise and focus on parameters' local importance, ignoring the rich topology information within DNNs. In this paper, we propose a novel multi-stage graph embedding technique based on graph neural networks (GNNs) to identify DNN topologies and use reinforcement learning (RL) to find a suitable compression policy. We performed resource-constrained (i.e., FLOPs) channel pruning and compared our approach with state-of-the-art model compression methods. We evaluated our method on various models from typical to mobile-friendly networks, such as ResNet family, VGG-16, MobileNet-v1/v2, and ShuffleNet. Results show that our method can achieve higher compression ratios with a minimal fine-tuning cost yet yields outstanding and competitive performance.
On-device machine learning is becoming a reality thanks to the availability of powerful hardware and model compression techniques. Typically, these models are pretrained on large GPU clusters and have enough parameters to generalise across a wide variety of inputs. In this work, we observe that a much smaller, personalised model can be employed to fit a specific scenario, resulting in both higher accuracy and faster execution. Nevertheless, on-device training is extremely challenging, imposing excessive computational and memory requirements even for flagship smartphones. At the same time, on-device data availability might be limited and samples are most frequently unlabelled. To this end, we introduce PersEPhonEE, a framework that attaches early exits on the model and personalises them on-device. These allow the model to progressively bypass a larger part of the computation as more personalised data become available. Moreover, we introduce an efficient on-device algorithm that trains the early exits in a semi-supervised manner at a fraction of the whole network's personalisation time. Results show that PersEPhonEE boosts accuracy by up to 15.9% while dropping the training cost by up to 2.2x and inference latency by 2.2-3.2x on average for the same accuracy, depending on the availability of labels on-device.
Channel pruning is formulated as a neural architecture search (NAS) problem recently. However, existing NAS-based methods are challenged by huge computational cost and inflexibility of applications. How to deal with multiple sparsity constraints simultaneously and speed up NAS-based channel pruning are still open challenges. In this paper, we propose a novel Accurate and Automatic Channel Pruning (AACP) method to address these problems. Firstly, AACP represents the structure of a model as a structure vector and introduces a pruning step vector to control the compressing granularity of each layer. Secondly, AACP utilizes Pruned Structure Accuracy Estimator (PSAE) to speed up the performance estimation process. Thirdly, AACP proposes Improved Differential Evolution (IDE) algorithm to search the optimal structure vector effectively. Because of IDE, AACP can deal with FLOPs constraint and model size constraint simultaneously and efficiently. Our method can be easily applied to various tasks and achieve state of the art performance. On CIFAR10, our method reduces $65\%$ FLOPs of ResNet110 with an improvement of $0.26\%$ top-1 accuracy. On ImageNet, we reduce $42\%$ FLOPs of ResNet50 with a small loss of $0.18\%$ top-1 accuracy and reduce $30\%$ FLOPs of MobileNetV2 with a small loss of $0.7\%$ top-1 accuracy. The source code will be released after publication.
Deep Convolutional Neural Networks (DCNNs) have shown promising performances in several visual recognition problems which motivated the researchers to propose popular architectures such as LeNet, AlexNet, VGGNet, ResNet, and many more. These architectures come at a cost of high computational complexity and parameter storage. To get rid of storage and computational complexity, deep model compression methods have been evolved. We propose a "History Based Filter Pruning (HBFP)" method that utilizes network training history for filter pruning. Specifically, we prune the redundant filters by observing similar patterns in the filter's L1-norms (absolute sum of weights) over the training epochs. We iteratively prune the redundant filters of a CNN in three steps. First, we train the model and select the filter pairs with redundant filters in each pair. Next, we optimize the network to ensure an increased measure of similarity between the filters in a pair. This optimization of the network facilitates us to prune one filter from each pair based on its importance without much information loss. Finally, we retrain the network to regain the performance, which is dropped due to filter pruning. We test our approach on popular architectures such as LeNet-5 on MNIST dataset; VGG-16, ResNet-56, and ResNet-110 on CIFAR-10 dataset, and ResNet-50 on ImageNet. The proposed pruning method outperforms the state-of-the-art in terms of FLOPs reduction (floating-point operations) by 97.98%, 83.42%, 78.43%, 74.95%, and 75.45% for LeNet-5, VGG-16, ResNet-56, ResNet-110, and ResNet-50, respectively, while maintaining the less error rate.
The performance of federated learning systems is bottlenecked by communication costs and training variance. The communication overhead problem is usually addressed by three communication-reduction techniques, namely, model compression, partial device participation, and periodic aggregation, at the cost of increased training variance. Different from traditional distributed learning systems, federated learning suffers from data heterogeneity (since the devices sample their data from possibly different distributions), which induces additional variance among devices during training. Various variance-reduced training algorithms have been introduced to combat the effects of data heterogeneity, while they usually cost additional communication resources to deliver necessary control information. Additionally, data privacy remains a critical issue in FL, and thus there have been attempts at bringing Differential Privacy to this framework as a mediator between utility and privacy requirements. This paper investigates the trade-offs between communication costs and training variance under a resource-constrained federated system theoretically and experimentally, and how communication reduction techniques interplay in a differentially private setting. The results provide important insights into designing practical privacy-aware federated learning systems.
There are many deep learning (e.g., DNN) powered mobile and wearable applications today continuously and unobtrusively sensing the ambient surroundings to enhance all aspects of human lives. To enable robust and private mobile sensing, DNN tends to be deployed locally on the resource-constrained mobile devices via model compression. The current practice either hand-crafted DNN compression techniques, i.e., for optimizing DNN-relative performance (e.g., parameter size), or on-demand DNN compression methods, i.e., for optimizing hardware-dependent metrics (e.g., latency), cannot be locally online because they require offline retraining to ensure accuracy. Also, none of them have correlated their efforts with runtime adaptive compression to consider the dynamic nature of the deployment context of mobile applications. To address those challenges, we present AdaSpring, a context-adaptive and self-evolutionary DNN compression framework. It enables the runtime adaptive DNN compression locally online. Specifically, it presents the ensemble training of a retraining-free and self-evolutionary network to integrate multiple alternative DNN compression configurations (i.e., compressed architectures and weights). It then introduces the runtime search strategy to quickly search for the most suitable compression configurations and evolve the corresponding weights. With evaluation on five tasks across three platforms and a real-world case study, experiment outcomes show that AdaSpring obtains up to 3.1x latency reduction, 4.2 x energy efficiency improvement in DNNs, compared to hand-crafted compression techniques, while only incurring <= 6.2ms runtime-evolution latency.
Knowledge distillation (KD), as an efficient and effective model compression technique, has been receiving considerable attention in deep learning. The key to its success is to transfer knowledge from a large teacher network to a small student one. However, most of the existing knowledge distillation methods consider only one type of knowledge learned from either instance features or instance relations via a specific distillation strategy in teacher-student learning. There are few works that explore the idea of transferring different types of knowledge with different distillation strategies in a unified framework. Moreover, the frequently used offline distillation suffers from a limited learning capacity due to the fixed teacher-student architecture. In this paper we propose a collaborative teacher-student learning via multiple knowledge transfer (CTSL-MKT) that prompts both self-learning and collaborative learning. It allows multiple students learn knowledge from both individual instances and instance relations in a collaborative way. While learning from themselves with self-distillation, they can also guide each other via online distillation. The experiments and ablation studies on four image datasets demonstrate that the proposed CTSL-MKT significantly outperforms the state-of-the-art KD methods.
Recent improvements in the predictive quality of natural language processing systems are often dependent on a substantial increase in the number of model parameters. This has led to various attempts of compressing such models, but existing methods have not considered the differences in the predictive power of various model components or in the generalizability of the compressed models. To understand the connection between model compression and out-of-distribution generalization, we define the task of compressing language representation models such that they perform best in a domain adaptation setting. We choose to address this problem from a causal perspective, attempting to estimate the average treatment effect (ATE) of a model component, such as a single layer, on the model's predictions. Our proposed ATE-guided Model Compression scheme (AMoC), generates many model candidates, differing by the model components that were removed. Then, we select the best candidate through a stepwise regression model that utilizes the ATE to predict the expected performance on the target domain. AMoC outperforms strong baselines on dozens of domain pairs across three text classification and sequence tagging tasks.
Artificial neural network has achieved the state-of-art performance in fault detection on the Tennessee Eastman process, but it often requires enormous memory to fund its massive parameters. In order to implement online real-time fault detection, three deep compression techniques (pruning, clustering, and quantization) are applied to reduce the computational burden. We have extensively studied 7 different combinations of compression techniques, all methods achieve high model compression rates over 64% while maintain high fault detection accuracy. The best result is applying all three techniques, which reduces the model sizes by 91.5% and remains a high accuracy over 94%. This result leads to a smaller storage requirement in production environments, and makes the deployment smoother in real world.
On-device Deep Neural Networks (DNNs) have recently gained more attention due to the increasing computing power of the mobile devices and the number of applications in Computer Vision (CV), Natural Language Processing (NLP), and Internet of Things (IoTs). Unfortunately, the existing efficient convolutional neural network (CNN) architectures designed for CV tasks are not directly applicable to NLP tasks and the tiny Recurrent Neural Network (RNN) architectures have been designed primarily for IoT applications. In NLP applications, although model compression has seen initial success in on-device text classification, there are at least three major challenges yet to be addressed: adversarial robustness, explainability, and personalization. Here we attempt to tackle these challenges by designing a new training scheme for model compression and adversarial robustness, including the optimization of an explainable feature mapping objective, a knowledge distillation objective, and an adversarially robustness objective. The resulting compressed model is personalized using on-device private training data via fine-tuning. We perform extensive experiments to compare our approach with both compact RNN (e.g., FastGRNN) and compressed RNN (e.g., PRADO) architectures in both natural and adversarial NLP test settings.
We introduce an automated tool for deploying ultra low-latency, low-power deep neural networks with convolutional layers on FPGAs. By extending the hls4ml library, we demonstrate an inference latency of $5\,\mu$s using convolutional architectures, targeting microsecond latency applications like those at the CERN Large Hadron Collider. Considering benchmark models trained on the Street View House Numbers Dataset, we demonstrate various methods for model compression in order to fit the computational constraints of a typical FPGA device used in trigger and data acquisition systems of particle detectors. In particular, we discuss pruning and quantization-aware training, and demonstrate how resource utilization can be significantly reduced with little to no loss in model accuracy. We show that the FPGA critical resource consumption can be reduced by 97% with zero loss in model accuracy, and by 99% when tolerating a 6% accuracy degradation.
Network pruning and quantization are proven to be effective ways for deep model compression. To obtain a highly compact model, most methods first perform network pruning and then conduct network quantization based on the pruned model. However, this strategy may ignore that they would affect each other and thus performing them separately may lead to sub-optimal performance. To address this, performing pruning and quantization jointly is essential. Nevertheless, how to make a trade-off between pruning and quantization is non-trivial. Moreover, existing compression methods often rely on some pre-defined compression configurations. Some attempts have been made to search for optimal configurations, which however may take unbearable optimization cost. To address the above issues, we devise a simple yet effective method named Single-path Bit Sharing (SBS). Specifically, we first consider network pruning as a special case of quantization, which provides a unified view for pruning and quantization. We then introduce a single-path model to encode all candidate compression configurations. In this way, the configuration search problem is transformed into a subset selection problem, which significantly reduces the number of parameters, computational cost and optimization difficulty. Relying on the single-path model, we further introduce learnable binary gates to encode the choice of bitwidth. By jointly training the binary gates in conjunction with network parameters, the compression configurations of each layer can be automatically determined. Extensive experiments on both CIFAR-100 and ImageNet show that SBS is able to significantly reduce computational cost while achieving promising performance. For example, our SBS compressed MobileNetV2 achieves 22.6x Bit-Operation (BOP) reduction with only 0.1% drop in the Top-1 accuracy.
As neural networks gain widespread adoption in embedded devices, there is a need for model compression techniques to facilitate deployment in resource-constrained environments. Quantization is one of the go-to methods yielding state-of-the-art model compression. Most approaches take a fully trained model, apply different heuristics to determine the optimal bit-precision for different layers of the network, and retrain the network to regain any drop in accuracy. Based on Activation Density (AD)-the proportion of non-zero activations in a layer-we propose an in-training quantization method. Our method calculates bit-width for each layer during training yielding a mixed precision model with competitive accuracy. Since we train lower precision models during training, our approach yields the final quantized model at lower training complexity and also eliminates the need for re-training. We run experiments on benchmark datasets like CIFAR-10, CIFAR-100, TinyImagenet on VGG19/ResNet18 architectures and report the accuracy and energy estimates for the same. We achieve ~4.5x benefit in terms of estimated multiply-and-accumulate (MAC) reduction while reducing the training complexity by 50% in our experiments. To further evaluate the energy benefits of our proposed method, we develop a mixed-precision scalable Process In Memory (PIM) hardware accelerator platform. The hardware platform incorporates shift-add functionality for handling multi-bit precision neural network models. Evaluating the quantized models obtained with our proposed method on the PIM platform yields ~5x energy reduction compared to 16-bit models. Additionally, we find that integrating AD based quantization with AD based pruning (both conducted during training) yields up to ~198x and ~44x energy reductions for VGG19 and ResNet18 architectures respectively on PIM platform compared to baseline 16-bit precision, unpruned models.
From small screenshots to large videos, documents take up a bulk of space in a modern smartphone. Documents in a phone can accumulate from various sources, and with the high storage capacity of mobiles, hundreds of documents are accumulated in a short period. However, searching or managing documents remains an onerous task, since most search methods depend on meta-information or only text in a document. In this paper, we showcase that a single modality is insufficient for classification and present a novel pipeline to classify documents on-device, thus preventing any private user data transfer to server. For this task, we integrate an open-source library for Optical Character Recognition (OCR) and our novel model architecture in the pipeline. We optimise the model for size, a necessary metric for on-device inference. We benchmark our classification model with a standard multimodal dataset FOOD-101 and showcase competitive results with the previous State of the Art with 30% model compression.
Heavily overparameterized language models such as BERT, XLNet and T5 have achieved impressive success in many NLP tasks. However, their high model complexity requires enormous computation resources and extremely long training time for both pre-training and fine-tuning. Many works have studied model compression on large NLP models, but only focusing on reducing inference time while still requiring an expensive training process. Other works use extremely large batch sizes to shorten the pre-training time, at the expense of higher computational resource demands. In this paper, inspired by the Early-Bird Lottery Tickets recently studied for computer vision tasks, we propose EarlyBERT, a general computationally-efficient training algorithm applicable to both pre-training and fine-tuning of large-scale language models. By slimming the self-attention and fully-connected sub-layers inside a transformer, we are the first to identify structured winning tickets in the early stage of BERT training. We apply those tickets towards efficient BERT training, and conduct comprehensive pre-training and fine-tuning experiments on GLUE and SQuAD downstream tasks. Our results show that EarlyBERT achieves comparable performance to standard BERT, with 35~45% less training time. Code is available at https://github.com/VITA-Group/EarlyBERT.
The rapid development of large pre-trained language models has greatly increased the demand for model compression techniques, among which quantization is a popular solution. In this paper, we propose BinaryBERT, which pushes BERT quantization to the limit by weight binarization. We find that a binary BERT is hard to be trained directly than a ternary counterpart due to its complex and irregular loss landscape. Therefore, we propose ternary weight splitting, which initializes BinaryBERT by equivalently splitting from a half-sized ternary network. The binary model thus inherits the good performance of the ternary one, and can be further enhanced by fine-tuning the new architecture after splitting. Empirical results show that our BinaryBERT has only a slight performance drop compared with the full-precision model while being 24x smaller, achieving the state-of-the-art compression results on the GLUE and SQuAD benchmarks.
Knowledge Distillation (KD) is a common knowledge transfer algorithm used for model compression across a variety of deep learning based natural language processing (NLP) solutions. In its regular manifestations, KD requires access to the teacher's training data for knowledge transfer to the student network. However, privacy concerns, data regulations and proprietary reasons may prevent access to such data. We present, to the best of our knowledge, the first work on Zero-Shot Knowledge Distillation for NLP, where the student learns from the much larger teacher without any task specific data. Our solution combines out of domain data and adversarial training to learn the teacher's output distribution. We investigate six tasks from the GLUE benchmark and demonstrate that we can achieve between 75% and 92% of the teacher's classification score (accuracy or F1) while compressing the model 30 times.
Network pruning is a widely used technique to reduce computation cost and model size for deep neural networks. However, the typical three-stage pipeline, i.e., training, pruning and retraining (fine-tuning) significantly increases the overall training trails. In this paper, we develop a systematic weight-pruning optimization approach based on Surrogate Lagrangian relaxation (SLR), which is tailored to overcome difficulties caused by the discrete nature of the weight-pruning problem while ensuring fast convergence. We further accelerate the convergence of the SLR by using quadratic penalties. Model parameters obtained by SLR during the training phase are much closer to their optimal values as compared to those obtained by other state-of-the-art methods. We evaluate the proposed method on image classification tasks, i.e., ResNet-18 and ResNet-50 using ImageNet, and ResNet-18, ResNet-50 and VGG-16 using CIFAR-10, as well as object detection tasks, i.e., YOLOv3 and YOLOv3-tiny using COCO 2014 and Ultra-Fast-Lane-Detection using TuSimple lane detection dataset. Experimental results demonstrate that our SLR-based weight-pruning optimization approach achieves higher compression rate than state-of-the-arts under the same accuracy requirement. It also achieves a high model accuracy even at the hard-pruning stage without retraining (reduces the traditional three-stage pruning to two-stage). Given a limited budget of retraining epochs, our approach quickly recovers the model accuracy.