In seeking for sparse and efficient neural network models, many previous works investigated on enforcing L1 or L0 regularizers to encourage weight sparsity during training. The L0 regularizer measures the parameter sparsity directly and is invariant to the scaling of parameter values, but it cannot provide useful gradients, and therefore requires complex optimization techniques. The L1 regularizer is almost everywhere differentiable and can be easily optimized with gradient descent. Yet it is not scale-invariant, causing the same shrinking rate to all parameters, which is inefficient in increasing sparsity. Inspired by the Hoyer measure (the ratio between L1 and L2 norms) used in traditional compressed sensing problems, we present DeepHoyer, a set of sparsity-inducing regularizers that are both differentiable almost everywhere and scale-invariant. Our experiments show that enforcing DeepHoyer regularizers can produce even sparser neural network models than previous works, under the same accuracy level. We also show that DeepHoyer can be applied to both element-wise and structural pruning.
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.
With rapidly scaling up of deep neural networks (DNNs), extensive research studies on network model compression such as weight pruning have been performed for improving deployment efficiency. This work aims to advance the compression beyond the weights to neuron activations. We propose an end-to-end Joint Pruning (JP) technique which integrates the activation pruning with the weight pruning. By distinguishing and taking on the different significance of neuron responses and connections during learning, the generated network, namely JPnet, optimizes the sparsity of activations and weights for improving execution efficiency. To our best knowledge, JP is the first technique that simultaneously explores the redundancy in both weights and activations. The derived deep sparsification in the JPnet reveals more optimizing potentialities for the existing DNN accelerators dedicated for sparse matrix operations. The effectiveness of JP technique is thoroughly evaluated through various network models with different activation functions and on different datasets. With $<0.4\%$ degradation on testing accuracy, a JPnet can save $71.1\% \sim 96.35\%$ of computation cost, compared to the original dense models with up to $5.8\times$ and $10\times$ reductions in activation and weight numbers, respectively. Compared to state-of-the-art weight pruning technique, JPnet can further reduce the computation cost $1.2\times \sim 2.7\times$.
We propose AutoGrow to automate depth discovery in Deep Neural Networks (DNNs): starting from a shallow seed architecture, AutoGrow grows new layers if the growth improves the accuracy; otherwise, the growth stops and the network depth is discovered. The residual and plain blocks are used as growing sub-modules to study DNNs with and without shortcuts. We propose generic growing and stopping policies to minimize human efforts spent on the optimal depth search. Our experiments show that by applying the same policy to different tasks, AutoGrow can always discover network depth effectively and achieve state-of-the-art accuracy on various datasets of MNIST, FashionMNIST, SVHN, CIFAR10, CIFAR100 and ImageNet. Comparing to Neural Architecture Search (NAS) that often designs a gigantic search space and consumes tremendous resources, AutoGrow lies at the other end of the research spectrum: it focuses on efficient depth discovery and reduces the growing and searching time to a level similar to that of training a single DNN. Thus, AutoGrow is able to scale up to large datasets such as ImageNet. Our study also reveals that previous Network Morphism is sub-optimal for increasing layer depth. Finally, we demonstrate that AutoGrow enables the training of deeper plain networks, which has been problematic even using Batch Normalization.
Adversarial attacks have exposed a significant security vulnerability in state-of-the-art machine learning models. Among these models include deep reinforcement learning agents. The existing methods for attacking reinforcement learning agents assume the adversary either has access to the target agent's learned parameters or the environment that the agent interacts with. In this work, we propose a new class of threat models, called snooping threat models, that are unique to reinforcement learning. In these snooping threat models, the adversary does not have the ability to personally interact with the environment, and can only eavesdrop on the action and reward signals being exchanged between agent and environment. We show that adversaries operating in these highly constrained threat models can still launch devastating attacks against the target agent by training proxy models on related tasks and leveraging the transferability of adversarial examples.
With the rise of artificial intelligence in recent years, Deep Neural Networks (DNNs) have been widely used in many domains. To achieve high performance and energy efficiency, hardware acceleration (especially inference) of DNNs is intensively studied both in academia and industry. However, we still face two challenges: large DNN models and datasets, which incur frequent off-chip memory accesses; and the training of DNNs, which is not well-explored in recent accelerator designs. To truly provide high throughput and energy efficient acceleration for the training of deep and large models, we inevitably need to use multiple accelerators to explore the coarse-grain parallelism, compared to the fine-grain parallelism inside a layer considered in most of the existing architectures. It poses the key research question to seek the best organization of computation and dataflow among accelerators. In this paper, we propose a solution HyPar to determine layer-wise parallelism for deep neural network training with an array of DNN accelerators. HyPar partitions the feature map tensors (input and output), the kernel tensors, the gradient tensors, and the error tensors for the DNN accelerators. A partition constitutes the choice of parallelism for weighted layers. The optimization target is to search a partition that minimizes the total communication during training a complete DNN. To solve this problem, we propose a communication model to explain the source and amount of communications. Then, we use a hierarchical layer-wise dynamic programming method to search for the partition for each layer.
In recent years, deep neural networks demonstrated state-of-the-art performance in a large variety of tasks and therefore have been adopted in many applications. On the other hand, the latest studies revealed that neural networks are vulnerable to adversarial examples obtained by carefully adding small perturbation to legitimate samples. Based upon the observation, many attack methods were proposed. Among them, the optimization-based CW attack is the most powerful as the produced adversarial samples present much less distortion compared to other methods. The better attacking effect, however, comes at the cost of running more iterations and thus longer computation time to reach desirable results. In this work, we propose to leverage the information of gradients as a guidance during the search of adversaries. More specifically, directly incorporating the gradients into the perturbation can be regarded as a constraint added to the optimization process. We intuitively and empirically prove the rationality of our method in reducing the search space. Our experiments show that compared to the original CW attack, the proposed method requires fewer iterations towards adversarial samples, obtaining a higher success rate and resulting in smaller $\ell_2$ distortion.
The success of deep learning research has catapulted deep models into production systems that our society is becoming increasingly dependent on, especially in the image and video domains. However, recent work has shown that these largely uninterpretable models exhibit glaring security vulnerabilities in the presence of an adversary. In this work, we develop a powerful untargeted adversarial attack for action recognition systems in both white-box and black-box settings. Action recognition models differ from image-classification models in that their inputs contain a temporal dimension, which we explicitly target in the attack. Drawing inspiration from image classifier attacks, we create new attacks which achieve state-of-the-art success rates on a two-stream classifier trained on the UCF-101 dataset. We find that our attacks can significantly degrade a model's performance with sparsely and imperceptibly perturbed examples. We also demonstrate the transferability of our attacks to black-box action recognition systems.
Distributed learning systems have enabled training large-scale models over large amount of data in significantly shorter time. In this paper, we focus on decentralized distributed deep learning systems and aim to achieve differential privacy with good convergence rate and low communication cost. To achieve this goal, we propose a new learning algorithm LEASGD (Leader-Follower Elastic Averaging Stochastic Gradient Descent), which is driven by a novel Leader-Follower topology and a differential privacy model.We provide a theoretical analysis of the convergence rate and the trade-off between the performance and privacy in the private setting.The experimental results show that LEASGD outperforms state-of-the-art decentralized learning algorithm DPSGD by achieving steadily lower loss within the same iterations and by reducing the communication cost by 30%. In addition, LEASGD spends less differential privacy budget and has higher final accuracy result than DPSGD under private setting.
Neural networks have shown great performance in cognitive tasks. When deploying network models on mobile devices with limited resources, weight quantization has been widely adopted. Binary quantization obtains the highest compression but usually results in big accuracy drop. In practice, 8-bit or 16-bit quantization is often used aiming at maintaining the same accuracy as the original 32-bit precision. We observe different layers have different accuracy sensitivity of quantization. Thus judiciously selecting different precision for different layers/structures can potentially produce more efficient models compared to traditional quantization methods by striking a better balance between accuracy and compression rate. In this work, we propose a fine-grained quantization approach for deep neural network compression by relaxing the search space of quantization bitwidth from discrete to a continuous domain. The proposed approach applies gradient descend based optimization to generate a mixed-precision quantization scheme that outperforms the accuracy of traditional quantization methods under the same compression rate.