Reducing the size of neural network models is a critical step in moving AI from a cloud-centric to an edge-centric (i.e. on-device) compute paradigm. This shift from cloud to edge is motivated by a number of factors including reduced latency, improved security, and higher flexibility of AI algorithms across several application domains (e.g. transportation, healthcare, defense, etc.). However, it is currently unclear how model compression techniques may affect the robustness of AI algorithms against adversarial attacks. This paper explores the effect of quantization, one of the most common compression techniques, on the adversarial robustness of neural networks. Specifically, we investigate and model the accuracy of quantized neural networks on adversarially-perturbed images. Results indicate that for simple gradient-based attacks, quantization can either improve or degrade adversarial robustness depending on the attack strength.
Recent applications pose requirements of both cross-domain knowledge transfer and model compression to machine learning models due to insufficient training data and limited computational resources. In this paper, we propose a new knowledge distillation model, named Spirit Distillation (SD), which is a model compression method with multi-domain knowledge transfer. The compact student network mimics out a representation equivalent to the front part of the teacher network, through which the general knowledge can be transferred from the source domain (teacher) to the target domain (student). To further improve the robustness of the student, we extend SD to Enhanced Spirit Distillation (ESD) in exploiting a more comprehensive knowledge by introducing the proximity domain which is similar to the target domain for feature extraction. Results demonstrate that our method can boost mIOU and high-precision accuracy by 1.4% and 8.2% respectively with 78.2% segmentation variance, and can gain a precise compact network with only 41.8% FLOPs.
In a backdoor attack on a machine learning model, an adversary produces a model that performs well on normal inputs but outputs targeted misclassifications on inputs containing a small trigger pattern. Model compression is a widely-used approach for reducing the size of deep learning models without much accuracy loss, enabling resource-hungry models to be compressed for use on resource-constrained devices. In this paper, we study the risk that model compression could provide an opportunity for adversaries to inject stealthy backdoors. We design stealthy backdoor attacks such that the full-sized model released by adversaries appears to be free from backdoors (even when tested using state-of-the-art techniques), but when the model is compressed it exhibits highly effective backdoors. We show this can be done for two common model compression techniques -- model pruning and model quantization. Our findings demonstrate how an adversary may be able to hide a backdoor as a compression artifact, and show the importance of performing security tests on the models that will actually be deployed not their precompressed version.
Spiking Neural Networks (SNNs) are a promising alternative to traditional deep learning methods since they perform event-driven information processing. However, a major drawback of SNNs is high inference latency. The efficiency of SNNs could be enhanced using compression methods such as pruning and quantization. Notably, SNNs, unlike their non-spiking counterparts, consist of a temporal dimension, the compression of which can lead to latency reduction. In this paper, we propose spatial and temporal pruning of SNNs. First, structured spatial pruning is performed by determining the layer-wise significant dimensions using principal component analysis of the average accumulated membrane potential of the neurons. This step leads to 10-14X model compression. Additionally, it enables inference with lower latency and decreases the spike count per inference. To further reduce latency, temporal pruning is performed by gradually reducing the timesteps while training. The networks are trained using surrogate gradient descent based backpropagation and we validate the results on CIFAR10 and CIFAR100, using VGG architectures. The spatiotemporally pruned SNNs achieve 89.04% and 66.4% accuracy on CIFAR10 and CIFAR100, respectively, while performing inference with 3-30X reduced latency compared to state-of-the-art SNNs. Moreover, they require 8-14X lesser compute energy compared to their unpruned standard deep learning counterparts. The energy numbers are obtained by multiplying the number of operations with energy per operation. These SNNs also provide 1-4% higher robustness against Gaussian noise corrupted inputs. Furthermore, we perform weight quantization and find that performance remains reasonably stable up to 5-bit quantization.
We propose Skip-Convolutions to leverage the large amount of redundancies in video streams and save computations. Each video is represented as a series of changes across frames and network activations, denoted as residuals. We reformulate standard convolution to be efficiently computed on residual frames: each layer is coupled with a binary gate deciding whether a residual is important to the model prediction,~\eg foreground regions, or it can be safely skipped, e.g. background regions. These gates can either be implemented as an efficient network trained jointly with convolution kernels, or can simply skip the residuals based on their magnitude. Gating functions can also incorporate block-wise sparsity structures, as required for efficient implementation on hardware platforms. By replacing all convolutions with Skip-Convolutions in two state-of-the-art architectures, namely EfficientDet and HRNet, we reduce their computational cost consistently by a factor of 3~4x for two different tasks, without any accuracy drop. Extensive comparisons with existing model compression, as well as image and video efficiency methods demonstrate that Skip-Convolutions set a new state-of-the-art by effectively exploiting the temporal redundancies in videos.
A popular approach to model compression is to train an inexpensive student model to mimic the class probabilities of a highly accurate but cumbersome teacher model. Surprisingly, this two-step knowledge distillation process often leads to higher accuracy than training the student directly on labeled data. To explain and enhance this phenomenon, we cast knowledge distillation as a semiparametric inference problem with the optimal student model as the target, the unknown Bayes class probabilities as nuisance, and the teacher probabilities as a plug-in nuisance estimate. By adapting modern semiparametric tools, we derive new guarantees for the prediction error of standard distillation and develop two enhancements -- cross-fitting and loss correction -- to mitigate the impact of teacher overfitting and underfitting on student performance. We validate our findings empirically on both tabular and image data and observe consistent improvements from our knowledge distillation enhancements.
We propose DiffQ a differentiable method for model compression for quantizing model parameters without gradient approximations (e.g., Straight Through Estimator). We suggest adding independent pseudo quantization noise to model parameters during training to approximate the effect of a quantization operator. DiffQ is differentiable both with respect to the unquantized weights and the number of bits used. Given a single hyper-parameter balancing between the quantized model size and accuracy, DiffQ optimizes the number of bits used per individual weight or groups of weights, in end-to-end training. We experimentally verify that our method is competitive with STE based quantization techniques on several benchmarks and architectures for image classification, language modeling, and audio source separation. For instance, on the ImageNet dataset, DiffQ compresses a 12 layers transformer-based model by more than a factor of 8, (lower than 4 bits precision per weight on average), with a loss of 0.3% in model accuracy. Code is available at github.com/facebookresearch/diffq.
The concept of compressing deep Convolutional Neural Networks (CNNs) is essential to use limited computation, power, and memory resources on embedded devices. However, existing methods achieve this objective at the cost of a drop in inference accuracy in computer vision tasks. To address such a drawback, we propose a framework that leverages knowledge distillation along with customizable block-wise optimization to learn a lightweight CNN structure while preserving better control over the compression-performance tradeoff. Considering specific resource constraints, e.g., floating-point operations per inference (FLOPs) or model-parameters, our method results in a state of the art network compression while being capable of achieving better inference accuracy. In a comprehensive evaluation, we demonstrate that our method is effective, robust, and consistent with results over a variety of network architectures and datasets, at negligible training overhead. In particular, for the already compact network MobileNet_v2, our method offers up to 2x and 5.2x better model compression in terms of FLOPs and model-parameters, respectively, while getting 1.05% better model performance than the baseline network.
Deep neural networks have been the driving force behind the success in classification tasks, e.g., object and audio recognition. Impressive results and generalization have been achieved by a variety of recently proposed architectures, the majority of which are seemingly disconnected. In this work, we cast the study of deep classifiers under a unifying framework. In particular, we express state-of-the-art architectures (e.g., residual and non-local networks) in the form of different degree polynomials of the input. Our framework provides insights on the inductive biases of each model and enables natural extensions building upon their polynomial nature. The efficacy of the proposed models is evaluated on standard image and audio classification benchmarks. The expressivity of the proposed models is highlighted both in terms of increased model performance as well as model compression. Lastly, the extensions allowed by this taxonomy showcase benefits in the presence of limited data and long-tailed data distributions. We expect this taxonomy to provide links between existing domain-specific architectures. The source code is available at \url{https://github.com/grigorisg9gr/polynomials-for-augmenting-NNs}.
Significant memory and computational requirements of large deep neural networks restrict their application on edge devices. Knowledge distillation (KD) is a prominent model compression technique for deep neural networks in which the knowledge of a trained large teacher model is transferred to a smaller student model. The success of knowledge distillation is mainly attributed to its training objective function, which exploits the soft-target information (also known as "dark knowledge") besides the given regular hard labels in a training set. However, it is shown in the literature that the larger the gap between the teacher and the student networks, the more difficult is their training using knowledge distillation. To address this shortcoming, we propose an improved knowledge distillation method (called Annealing-KD) by feeding the rich information provided by the teacher's soft-targets incrementally and more efficiently. Our Annealing-KD technique is based on a gradual transition over annealed soft-targets generated by the teacher at different temperatures in an iterative process, and therefore, the student is trained to follow the annealed teacher output in a step-by-step manner. This paper includes theoretical and empirical evidence as well as practical experiments to support the effectiveness of our Annealing-KD method. We did a comprehensive set of experiments on different tasks such as image classification (CIFAR-10 and 100) and NLP language inference with BERT-based models on the GLUE benchmark and consistently got superior results.
This paper presents a computational solution that enables continuous cardiac monitoring through cross-modality inference of electrocardiogram (ECG). While some smartwatches now allow users to obtain a 30-second ECG test by tapping a built-in bio-sensor, these short-term ECG tests often miss intermittent and asymptomatic abnormalities of cardiac functions. It is also infeasible to expect persistently active user participation for long-term continuous cardiac monitoring in order to capture these and other types of cardiac abnormalities. To alleviate the need for continuous user attention and active participation, we design a lightweight neural network that infers ECG from the photoplethysmogram (PPG) signal sensed at the skin surface by a wearable optical sensor. We also develop a diagnosis-oriented training strategy to enable the neural network to capture the pathological features of ECG, aiming to increase the utility of reconstructed ECG signals for screening cardiovascular diseases (CVDs). We also leverage model interpretation to obtain insights from data-driven models, for example, to reveal some associations between CVDs and ECG/PPG and to demonstrate how the neural network copes with motion artifacts in the ambulatory application. The experimental results on three datasets demonstrate the feasibility of inferring ECG from PPG, achieving a high fidelity of ECG reconstruction with only about 40K parameters.
Implanted devices providing real-time neural activity classification and control are increasingly used to treat neurological disorders, such as epilepsy and Parkinson's disease. Classification performance is critical to identifying brain states appropriate for the therapeutic action. However, advanced algorithms that have shown promise in offline studies, in particular deep learning (DL) methods, have not been deployed on resource-restrained neural implants. Here, we designed and optimized three embedded DL models of commonly adopted architectures and evaluated their inference performance in a case study of seizure detection. A deep neural network (DNN), a convolutional neural network (CNN), and a long short-term memory (LSTM) network were designed to classify ictal, preictal, and interictal phases from the CHB-MIT scalp EEG database. After iterative model compression and quantization, the algorithms were deployed on a general-purpose, off-the-shelf microcontroller. Inference sensitivity, false positive rate, execution time, memory size, and power consumption were quantified. For seizure event detection, the sensitivity and FPR (h-1) for the DNN, CNN, and LSTM models were 87.36%/0.169, 96.70%/0.102, and 97.61%/0.071, respectively. Predicting seizures for early warnings was also feasible. The implemented compression and quantization achieved a significant saving of power and memory with an accuracy degradation of less than 0.5%. Edge DL models achieved performance comparable to many prior implementations that had no time or computational resource limitations. Generic microcontrollers can provide the required memory and computational resources, while model designs can be migrated to ASICs for further optimization. The results suggest that edge DL inference is a feasible option for future neural implants to improve classification performance and therapeutic outcomes.
Knowledge distillation has been widely used to produce portable and efficient neural networks which can be well applied on edge devices for computer vision tasks. However, almost all top-performing knowledge distillation methods need to access the original training data, which usually has a huge size and is often unavailable. To tackle this problem, we propose a novel data-free approach in this paper, named Dual Discriminator Adversarial Distillation (DDAD) to distill a neural network without any training data or meta-data. To be specific, we use a generator to create samples through dual discriminator adversarial distillation, which mimics the original training data. The generator not only uses the pre-trained teacher's intrinsic statistics in existing batch normalization layers but also obtains the maximum discrepancy from the student model. Then the generated samples are used to train the compact student network under the supervision of the teacher. The proposed method obtains an efficient student network which closely approximates its teacher network, despite using no original training data. Extensive experiments are conducted to to demonstrate the effectiveness of the proposed approach on CIFAR-10, CIFAR-100 and Caltech101 datasets for classification tasks. Moreover, we extend our method to semantic segmentation tasks on several public datasets such as CamVid and NYUv2. All experiments show that our method outperforms all baselines for data-free knowledge distillation.
Deep convolutional neural networks have made outstanding contributions in many fields such as computer vision in the past few years and many researchers published well-trained network for downloading. But recent studies have shown serious concerns about integrity due to model-reuse attacks and backdoor attacks. In order to protect these open-source networks, many algorithms have been proposed such as watermarking. However, these existing algorithms modify the contents of the network permanently and are not suitable for integrity authentication. In this paper, we propose a reversible watermarking algorithm for integrity authentication. Specifically, we present the reversible watermarking problem of deep convolutional neural networks and utilize the pruning theory of model compression technology to construct a host sequence used for embedding watermarking information by histogram shift. As shown in the experiments, the influence of embedding reversible watermarking on the classification performance is less than 0.5% and the parameters of the model can be fully recovered after extracting the watermarking. At the same time, the integrity of the model can be verified by applying the reversible watermarking: if the model is modified illegally, the authentication information generated by original model will be absolutely different from the extracted watermarking information.
As the complexity of Deep Neural Network (DNN) models increases, their deployment on mobile devices becomes increasingly challenging, especially in complex vision tasks such as image classification. Many of recent contributions aim either to produce compact models matching the limited computing capabilities of mobile devices or to offload the execution of such burdensome models to a compute-capable device at the network edge - the edge servers. In this paper, we propose to modify the structure and training process of DNN models for complex image classification tasks to achieve in-network compression in the early network layers. Our training process stems from knowledge distillation, a technique that has been traditionally used to build small - student - models mimicking the output of larger - teacher - models. Here, we adopt this idea to obtain aggressive compression while preserving accuracy. Our results demonstrate that our approach is effective for state-of-the-art models trained over complex datasets, and can extend the parameter region in which edge computing is a viable and advantageous option. Additionally, we demonstrate that in many settings of practical interest we reduce the inference time with respect to specialized models such as MobileNet v2 executed at the mobile device, while improving accuracy.
Offloading the execution of complex Deep Neural Networks (DNNs) models to compute-capable devices at the network edge, that is, edge servers, can significantly reduce capture-to-output delay. However, the communication link between the mobile devices and edge servers can become the bottleneck when channel conditions are poor. We propose a framework to split DNNs for image processing and minimize capture-to-output delay in a wide range of network conditions and computing parameters. The core idea is to split the DNN models into head and tail models, where the two sections are deployed at the mobile device and edge server, respectively. Different from prior literature presenting DNN splitting frameworks, we distill the architecture of the head DNN to reduce its computational complexity and introduce a bottleneck, thus minimizing processing load at the mobile device as well as the amount of wirelessly transferred data. Our results show 98% reduction in used bandwidth and 85% in computation load compared to straightforward splitting.
The unstructured sparsity after pruning poses a challenge to the efficient implementation of deep learning models in existing regular architectures like systolic arrays. On the other hand, coarse-grained structured pruning is suitable for implementation in regular architectures but tends to have higher accuracy loss than unstructured pruning when the pruned models are of the same size. In this work, we propose a model compression method based on a novel weight permutation scheme to fully exploit the fine-grained weight sparsity in the hardware design. Through permutation, the optimal arrangement of the weight matrix is obtained, and the sparse weight matrix is further compressed to a small and dense format to make full use of the hardware resources. Two pruning granularities are explored. In addition to the unstructured weight pruning, we also propose a more fine-grained subword-level pruning to further improve the compression performance. Compared to the state-of-the-art works, the matrix compression rate is significantly improved from 5.88x to 14.13x. As a result, the throughput and energy efficiency are improved by 2.75 and 1.86 times, respectively.
This work presents self-supervised learning methods for developing monaural speaker-specific (i.e., personalized) speech enhancement models. While generalist models must broadly address many speakers, specialist models can adapt their enhancement function towards a particular speaker's voice, expecting to solve a narrower problem. Hence, specialists are capable of achieving more optimal performance in addition to reducing computational complexity. However, naive personalization methods can require clean speech from the target user, which is inconvenient to acquire, e.g., due to subpar recording conditions. To this end, we pose personalization as either a zero-shot task, in which no additional clean speech of the target speaker is used for training, or a few-shot learning task, in which the goal is to minimize the duration of the clean speech used for transfer learning. With this paper, we propose self-supervised learning methods as a solution to both zero- and few-shot personalization tasks. The proposed methods are designed to learn the personalized speech features from unlabeled data (i.e., in-the-wild noisy recordings from the target user) without knowing the corresponding clean sources. Our experiments investigate three different self-supervised learning mechanisms. The results show that self-supervised models achieve zero-shot and few-shot personalization using fewer model parameters and less clean data from the target user, achieving the data efficiency and model compression goals.
The predictive advantage of combining several different predictive models is widely accepted. Particularly in time series forecasting problems, this combination is often dynamic to cope with potential non-stationary sources of variation present in the data. Despite their superior predictive performance, ensemble methods entail two main limitations: high computational costs and lack of transparency. These issues often preclude the deployment of such approaches, in favour of simpler yet more efficient and reliable ones. In this paper, we leverage the idea of model compression to address this problem in time series forecasting tasks. Model compression approaches have been mostly unexplored for forecasting. Their application in time series is challenging due to the evolving nature of the data. Further, while the literature focuses on neural networks, we apply model compression to distinct types of methods. In an extensive set of experiments, we show that compressing dynamic forecasting ensembles into an individual model leads to a comparable predictive performance and a drastic reduction in computational costs. Further, the compressed individual model with best average rank is a rule-based regression model. Thus, model compression also leads to benefits in terms of model interpretability. The experiments carried in this paper are fully reproducible.
Neural network deployment on low-cost embedded systems, hence on microcontrollers (MCUs), has recently been attracting more attention than ever. Since MCUs have limited memory capacity as well as limited compute-speed, it is critical that we employ model compression, which reduces both memory and compute-speed requirements. In this paper, we add model compression, specifically Deep Compression, and further optimize Unlu's earlier work on arXiv, which efficiently deploys PyTorch models on MCUs. First, we prune the weights in convolutional and fully connected layers. Secondly, the remaining weights and activations are quantized to 8-bit integers from 32-bit floating-point. Finally, forward pass functions are compressed using special data structures for sparse matrices, which store only nonzero weights (without impacting performance and accuracy). In the case of the LeNet-5 model, the memory footprint was reduced by 12.45x, and the inference speed was boosted by 2.57x.