Resource is an important constraint when deploying Deep Neural Networks (DNNs) on mobile and edge devices. Existing works commonly adopt the cell-based search approach, which limits the flexibility of network patterns in learned cell structures. Moreover, due to the topology-agnostic nature of existing works, including both cell-based and node-based approaches, the search process is time consuming and the performance of found architecture may be sub-optimal. To address these problems, we propose AutoShrink, a topology-aware Neural Architecture Search(NAS) for searching efficient building blocks of neural architectures. Our method is node-based and thus can learn flexible network patterns in cell structures within a topological search space. Directed Acyclic Graphs (DAGs) are used to abstract DNN architectures and progressively optimize the cell structure through edge shrinking. As the search space intrinsically reduces as the edges are progressively shrunk, AutoShrink explores more flexible search space with even less search time. We evaluate AutoShrink on image classification and language tasks by crafting ShrinkCNN and ShrinkRNN models. ShrinkCNN is able to achieve up to 48% parameter reduction and save 34% Multiply-Accumulates (MACs) on ImageNet-1K with comparable accuracy of state-of-the-art (SOTA) models. Specifically, both ShrinkCNN and ShrinkRNN are crafted within 1.5 GPU hours, which is 7.2x and 6.7x faster than the crafting time of SOTA CNN and RNN models, respectively.
Recently, deep neural networks (DNN) have been widely used in speaker recognition area. In order to achieve fast response time and high accuracy, the requirements for hardware resources increase rapidly. However, as the speaker recognition application is often implemented on mobile devices, it is necessary to maintain a low computational cost while keeping high accuracy in far-field condition. In this paper, we apply structural sparsification on time-delay neural networks (TDNN) to remove redundant structures and accelerate the execution. On our targeted hardware, our model can remove 60% of parameters and only slightly increasing equal error rate (EER) by 0.18% while our structural sparse model can achieve more than 2x speedup.
To accelerate DNNs inference, low-rank approximation has been widely adopted because of its solid theoretical rationale and efficient implementations. Several previous works attempted to directly approximate a pre-trained model by low-rank decomposition; however, small approximation errors in parameters can ripple over a large prediction loss. Apparently, it is not optimal to separate low-rank approximation from training. Unlike previous works, this paper integrates low rank approximation and regularization into the training process. We propose Trained Rank Pruning (TRP), which alternates between low rank approximation and training. TRP maintains the capacity of the original network while imposing low-rank constraints during training. A nuclear regularization optimized by stochastic sub-gradient descent is utilized to further promote low rank in TRP. Networks trained with TRP has a low-rank structure in nature, and is approximated with negligible performance loss, thus eliminating fine-tuning after low rank approximation. The proposed method is comprehensively evaluated on CIFAR-10 and ImageNet, outperforming previous compression counterparts using low rank approximation. Our code is available at: https://github.com/yuhuixu1993/Trained-Rank-Pruning.
To accelerate DNNs inference, low-rank approximation has been widely adopted because of its solid theoretical rationale and efficient implementations. Several previous works attempted to directly approximate a pre-trained model by low-rank decomposition; however, small approximation errors in parameters can ripple over a large prediction loss. Apparently, it is not optimal to separate low-rank approximation from training. Unlike previous works, this paper integrates low rank approximation and regularization into the training process. We propose Trained Rank Pruning (TRP), which alternates between low rank approximation and training. TRP maintains the capacity of the original network while imposing low-rank constraints during training. A nuclear regularization optimized by stochastic sub-gradient descent is utilized to further promote low rank in TRP. Networks trained with TRP has a low-rank structure in nature, and is approximated with negligible performance loss, thus eliminating fine-tuning after low rank approximation. The proposed method is comprehensively evaluated on CIFAR-10 and ImageNet, outperforming previous compression counterparts using low rank approximation. Our code is available at: https://github.com/yuhuixu1993/Trained-Rank-Pruning.
Generative adversarial network (GAN) has greatly improved the quality of unsupervised image generation. Previous GAN-based methods often require a large amount of high-quality training data while producing a small number (e.g., tens) of classes. This work aims to scale up GANs to thousands of classes meanwhile reducing the use of high-quality data in training. We propose an image generation method based on conditional transferring features, which can capture pixel-level semantic changes when transforming low-quality images into high-quality ones. Moreover, self-supervision learning is integrated into our GAN architecture to provide more label-free semantic supervisory information observed from the training data. As such, training our GAN architecture requires much fewer high-quality images with a small number of additional low-quality images. The experiments on CIFAR-10 and STL-10 show that even removing 30% high-quality images from the training set, our method can still outperform previous ones. The scalability on object classes has been experimentally validated: our method with 30% fewer high-quality images obtains the best quality in generating 1,000 ImageNet classes, as well as generating all 3,755 classes of CASIA-HWDB1.0 Chinese handwriting characters.
Privacy recently emerges as a severe concern in deep learning, that is, sensitive data must be prohibited from being shared with the third party during deep neural network development. In this paper, we propose Morphed Learning (MoLe), an efficient and secure scheme to deliver deep learning data. MoLe has two main components: data morphing and Augmented Convolutional (Aug-Conv) layer. Data morphing allows data providers to send morphed data without privacy information, while Aug-Conv layer helps deep learning developers to apply their networks on the morphed data without performance penalty. MoLe provides stronger security while introducing lower overhead compared to GAZELLE (USENIX Security 2018), which is another method with no performance penalty on the neural network. When using MoLe for VGG-16 network on CIFAR dataset, the computational overhead is only 9% and the data transmission overhead is 5.12%. As a comparison, GAZELLE has computational overhead of 10,000 times and data transmission overhead of 421,000 times. In this setting, the attack success rate of adversary is 7.9 x 10^{-90} for MoLe and 2.9 x 10^{-30} for GAZELLE, respectively.
Deep learning has been widely utilized in many computer vision applications and achieved remarkable commercial success. However, running deep learning models on mobile devices is generally challenging due to limitation of the available computing resources. It is common to let the users send their service requests to cloud servers that run the large-scale deep learning models to process. Sending the data associated with the service requests to the cloud, however, impose risks on the user data privacy. Some prior arts proposed sending the features extracted from raw data (e.g., images) to the cloud. Unfortunately, these extracted features can still be exploited by attackers to recover raw images and to infer embedded private attributes (e.g., age, gender, etc.). In this paper, we propose an adversarial training framework DeepObfuscator that can prevent extracted features from being utilized to reconstruct raw images and infer private attributes, while retaining the useful information for the intended cloud service (i.e., image classification). DeepObfuscator includes a learnable encoder, namely, obfuscator that is designed to hide privacy-related sensitive information from the features by performingour proposed adversarial training algorithm. Our experiments on CelebAdataset show that the quality of the reconstructed images fromthe obfuscated features of the raw image is dramatically decreased from 0.9458 to 0.3175 in terms of multi-scale structural similarity (MS-SSIM). The person in the reconstructed image, hence, becomes hardly to be re-identified. The classification accuracy of the inferred private attributes that can be achieved by the attacker drops down to a random-guessing level, e.g., the accuracy of gender is reduced from 97.36% to 58.85%. As a comparison, the accuracy of the intended classification tasks performed via the cloud service drops by only 2%
Gradients to activations get involved in most of the calculations during back propagation procedure of Convolution Neural Networks (CNNs) training. However, an important known observation is that the majority of these gradients are close to zero, imposing little impact on weights update. These gradients can be then pruned to achieve high gradient sparsity during CNNs training and reduce the computational cost. In particular, we randomly change a gradient to zero or a threshold value if the gradient is below the threshold which is determined by the statistical distribution of activation gradients. We also theoretically proved that the training convergence of the CNN model can be guaranteed when the above activation gradient sparsification method is applied. We evaluated our method on AlexNet, MobileNet, ResNet-{18, 34, 50, 101, 152} with CIFAR-{10, 100} and ImageNet datasets. Experimental results show that our method can substantially reduce the computational cost with negligible accuracy loss or even accuracy improvement. Finally, we analyzed the benefits that the sparsity of activation gradients introduced in detail.
Deconvolution has been widespread in neural networks. For example, it is essential for performing unsupervised learning in generative adversarial networks or constructing fully convolutional networks for semantic segmentation. Resistive RAM (ReRAM)-based processing-in-memory architecture has been widely explored in accelerating convolutional computation and demonstrates good performance. Performing deconvolution on existing ReRAM-based accelerator designs, however, suffers from long latency and high energy consumption because deconvolutional computation includes not only convolution but also extra add-on operations. To realize the more efficient execution for deconvolution, we analyze its computation requirement and propose a ReRAM-based accelerator design, namely, RED. More specific, RED integrates two orthogonal methods, the pixel-wise mapping scheme for reducing redundancy caused by zero-inserting operations and the zero-skipping data flow for increasing the computation parallelism and therefore improving performance. Experimental evaluations show that compared to the state-of-the-art ReRAM-based accelerator, RED can speed up operation 3.69x~1.15x and reduce 8%~88.36% energy consumption.
Designing neural architectures for edge devices is subject to constraints of accuracy, inference latency, and computational cost. Traditionally, researchers manually craft deep neural networks to meet the needs of mobile devices. Neural Architecture Search (NAS) was proposed to automate the neural architecture design without requiring extensive domain expertise and significant manual efforts. Recent works utilized NAS to design mobile models by taking into account hardware constraints and achieved state-of-the-art accuracy with fewer parameters and less computational cost measured in Multiply-accumulates (MACs). To find highly compact neural architectures, existing works relies on predefined cells and directly applying width multiplier, which may potentially limit the model flexibility, reduce the useful feature map information, and cause accuracy drop. To conquer this issue, we propose GRAM(GRAph propagation as Meta-knowledge) that adopts fine-grained (node-wise) search method and accumulates the knowledge learned in updates into a meta-graph. As a result, GRAM can enable more flexible search space and achieve higher search efficiency. Without the constraints of predefined cell or blocks, we propose a new structure-level pruning method to remove redundant operations in neural architectures. SwiftNet, which is a set of models discovered by GRAM, outperforms MobileNet-V2 by 2.15x higher accuracy density and 2.42x faster with similar accuracy. Compared with FBNet, SwiftNet reduces the search cost by 26x and achieves 2.35x higher accuracy density and 1.47x speedup while preserving similar accuracy. SwiftNetcan obtain 63.28% top-1 accuracy on ImageNet-1K with only 53M MACs and 2.07M parameters. The corresponding inference latency is only 19.09 ms on Google Pixel 1.