Adversarial robustness is essential for security and reliability of machine learning systems. However, the adversarial robustness gained by sophisticated defense algorithms is easily erased as the neural network evolves to learn new tasks. This vulnerability can be addressed by fostering a novel capability for neural networks, termed continual robust learning, which focuses on both the (classification) performance and adversarial robustness on previous tasks during continuous learning. To achieve continuous robust learning, we propose an approach called Double Gradient Projection that projects the gradients for weight updates orthogonally onto two crucial subspaces -- one for stabilizing the smoothed sample gradients and another for stabilizing the final outputs of the neural network. The experimental results on four benchmarks demonstrate that the proposed approach effectively maintains continuous robustness against strong adversarial attacks, outperforming the baselines formed by combining the existing defense strategies and continual learning methods.
Decentralized learning has emerged as an alternative method to the popular parameter-server framework which suffers from high communication burden, single-point failure and scalability issues due to the need of a central server. However, most existing works focus on a single shared model for all workers regardless of the data heterogeneity problem, rendering the resulting model performing poorly on individual workers. In this work, we propose a novel personalized decentralized learning algorithm named DePRL via shared representations. Our algorithm relies on ideas from representation learning theory to learn a low-dimensional global representation collaboratively among all workers in a fully decentralized manner, and a user-specific low-dimensional local head leading to a personalized solution for each worker. We show that DePRL achieves, for the first time, a provable linear speedup for convergence with general non-linear representations (i.e., the convergence rate is improved linearly with respect to the number of workers). Experimental results support our theoretical findings showing the superiority of our method in data heterogeneous environments.
With the increasing demand for large-scale training of machine learning models, fully decentralized optimization methods have recently been advocated as alternatives to the popular parameter server framework. In this paradigm, each worker maintains a local estimate of the optimal parameter vector, and iteratively updates it by waiting and averaging all estimates obtained from its neighbors, and then corrects it on the basis of its local dataset. However, the synchronization phase is sensitive to stragglers. An efficient way to mitigate this effect is to consider asynchronous updates, where each worker computes stochastic gradients and communicates with other workers at its own pace. Unfortunately, fully asynchronous updates suffer from staleness of the stragglers' parameters. To address these limitations, we propose a fully decentralized algorithm DSGD-AAU with adaptive asynchronous updates via adaptively determining the number of neighbor workers for each worker to communicate with. We show that DSGD-AAU achieves a linear speedup for convergence (i.e., convergence performance increases linearly with respect to the number of workers). Experimental results on a suite of datasets and deep neural network models are provided to verify our theoretical results.
Identifying causality is a challenging task in many data-intensive scenarios. Many algorithms have been proposed for this critical task. However, most of them consider the learning algorithms for directed acyclic graph (DAG) of Bayesian network (BN). These BN-based models only have limited causal explainability because of the issue of Markov equivalence class. Moreover, they are dependent on the assumption of stationarity, whereas many sampling time series from complex system are nonstationary. The nonstationary time series bring dataset shift problem, which leads to the unsatisfactory performances of these algorithms. To fill these gaps, a novel causation model named Unique Causal Network (UCN) is proposed in this paper. Different from the previous BN-based models, UCN considers the influence of time delay, and proves the uniqueness of obtained network structure, which addresses the issue of Markov equivalence class. Furthermore, based on the decomposability property of UCN, a higher-order causal entropy (HCE) algorithm is designed to identify the structure of UCN in a distributed way. HCE algorithm measures the strength of causality by using nearest-neighbors entropy estimator, which works well on nonstationary time series. Finally, lots of experiments validate that HCE algorithm achieves state-of-the-art accuracy when time series are nonstationary, compared to the other baseline algorithms.
Consider a processor having access only to meta-data consisting of the timings of data packets and acknowledgment (ACK) packets from all nodes in a network. The meta-data report the source node of each packet, but not the destination nodes or the contents of the packets. The goal of the processor is to infer the network topology based solely on such information. Prior work leveraged causality metrics to identify which links are active. If the data timings and ACK timings of two nodes -- say node 1 and node 2, respectively -- are causally related, this may be taken as evidence that node 1 is communicating to node 2 (which sends back ACK packets to node 1). This paper starts with the observation that packet losses can weaken the causality relationship between data and ACK timing streams. To obviate this problem, a new Expectation Maximization (EM)-based algorithm is introduced -- EM-causality discovery algorithm (EM-CDA) -- which treats packet losses as latent variables. EM-CDA iterates between the estimation of packet losses and the evaluation of causality metrics. The method is validated through extensive experiments in wireless sensor networks on the NS-3 simulation platform.
The video super-resolution (VSR) method based on the recurrent convolutional network has strong temporal modeling capability for video sequences. However, the input information received by different recurrent units in the unidirectional recurrent convolutional network is unbalanced. Early reconstruction frames receive less temporal information, resulting in fuzzy or artifact results. Although the bidirectional recurrent convolution network can alleviate this problem, it greatly increases reconstruction time and computational complexity. It is also not suitable for many application scenarios, such as online super-resolution. To solve the above problems, we propose an end-to-end information prebuilt recurrent reconstruction network (IPRRN), consisting of an information prebuilt network (IPNet) and a recurrent reconstruction network (RRNet). By integrating sufficient information from the front of the video to build the hidden state needed for the initially recurrent unit to help restore the earlier frames, the information prebuilt network balances the input information difference before and after without backward propagation. In addition, we demonstrate a compact recurrent reconstruction network, which has significant improvements in recovery quality and time efficiency. Many experiments have verified the effectiveness of our proposed network, and compared with the existing state-of-the-art methods, our method can effectively achieve higher quantitative and qualitative evaluation performance.
Federated learning (FL) is a popular technique to train machine learning (ML) models with decentralized data. Extensive works have studied the performance of the global model; however, it is still unclear how the training process affects the final test accuracy. Exacerbating this problem is the fact that FL executions differ significantly from traditional ML with heterogeneous data characteristics across clients, involving more hyperparameters. In this work, we show that the final test accuracy of FL is dramatically affected by the early phase of the training process, i.e., FL exhibits critical learning periods, in which small gradient errors can have irrecoverable impact on the final test accuracy. To further explain this phenomenon, we generalize the trace of the Fisher Information Matrix (FIM) to FL and define a new notion called FedFIM, a quantity reflecting the local curvature of each clients from the beginning of the training in FL. Our findings suggest that the {\em initial learning phase} plays a critical role in understanding the FL performance. This is in contrast to many existing works which generally do not connect the final accuracy of FL to the early phase training. Finally, seizing critical learning periods in FL is of independent interest and could be useful for other problems such as the choices of hyperparameters such as the number of client selected per round, batch size, and more, so as to improve the performance of FL training and testing.
With the increasing demand for large-scale training of machine learning models, consensus-based distributed optimization methods have recently been advocated as alternatives to the popular parameter server framework. In this paradigm, each worker maintains a local estimate of the optimal parameter vector, and iteratively updates it by waiting and averaging all estimates obtained from its neighbors, and then corrects it on the basis of its local dataset. However, the synchronization phase can be time consuming due to the need to wait for \textit{stragglers}, i.e., slower workers. An efficient way to mitigate this effect is to let each worker wait only for updates from the fastest neighbors before updating its local parameter. The remaining neighbors are called \textit{backup workers.} To minimize the globally training time over the network, we propose a fully distributed algorithm to dynamically determine the number of backup workers for each worker. We show that our algorithm achieves a linear speedup for convergence (i.e., convergence performance increases linearly with respect to the number of workers). We conduct extensive experiments on MNIST and CIFAR-10 to verify our theoretical results.
The increasing accessibility of data provides substantial opportunities for understanding user behaviors. Unearthing anomalies in user behaviors is of particular importance as it helps signal harmful incidents such as network intrusions, terrorist activities, and financial frauds. Many visual analytics methods have been proposed to help understand user behavior-related data in various application domains. In this work, we survey the state of art in visual analytics of anomalous user behaviors and classify them into four categories including social interaction, travel, network communication, and transaction. We further examine the research works in each category in terms of data types, anomaly detection techniques, and visualization techniques, and interaction methods. Finally, we discuss the findings and potential research directions.
PSO is a widely recognized optimization algorithm inspired by social swarm. In this brief we present a heterogeneous strategy particle swarm optimization (HSPSO), in which a proportion of particles adopt a fully informed strategy to enhance the converging speed while the rest are singly informed to maintain the diversity. Our extensive numerical experiments show that HSPSO algorithm is able to obtain satisfactory solutions, outperforming both PSO and the fully informed PSO. The evolution process is examined from both structural and microscopic points of view. We find that the cooperation between two types of particles can facilitate a good balance between exploration and exploitation, yielding better performance. We demonstrate the applicability of HSPSO on the filter design problem.