Network pruning has been known to produce compact models without much accuracy degradation. However, how the pruning process affects a network's robustness and the working mechanism behind remain unresolved. In this work, we theoretically prove that the sparsity of network weights is closely associated with model robustness. Through experiments on a variety of adversarial pruning methods, we find that weights sparsity will not hurt but improve robustness, where both weights inheritance from the lottery ticket and adversarial training improve model robustness in network pruning. Based on these findings, we propose a novel adversarial training method called inverse weights inheritance, which imposes sparse weights distribution on a large network by inheriting weights from a small network, thereby improving the robustness of the large network.

The history of deep learning has shown that human-designed problem-specific networks can greatly improve the classification performance of general neural models. In most practical cases, however, choosing the optimal architecture for a given task remains a challenging problem. Recent architecture-search methods are able to automatically build neural models with strong performance but fail to fully appreciate the interaction between neural architecture and weights. This work investigates the problem of disentangling the role of the neural structure and its edge weights, by showing that well-trained architectures may not need any link-specific fine-tuning of the weights. We compare the performance of such weight-free networks (in our case these are binary networks with {0, 1}-valued weights) with random, weight-agnostic, pruned and standard fully connected networks. To find the optimal weight-agnostic network, we use a novel and computationally efficient method that translates the hard architecture-search problem into a feasible optimization problem.More specifically, we look at the optimal task-specific architectures as the optimal configuration of binary networks with {0, 1}-valued weights, which can be found through an approximate gradient descent strategy. Theoretical convergence guarantees of the proposed algorithm are obtained by bounding the error in the gradient approximation and its practical performance is evaluated on two real-world data sets. For measuring the structural similarities between different architectures, we use a novel spectral approach that allows us to underline the intrinsic differences between real-valued networks and weight-free architectures.

Adversarial examples of deep neural networks have been actively investigated on image-based classification, segmentation and detection tasks. However, adversarial robustness of video models still lacks exploration. While several studies have proposed how to generate adversarial videos, only a handful of approaches pertaining to the defense strategies have been published in the literature. Furthermore, these defense methods are limited to a single perturbation type and often fail to provide robustness to Lp-bounded attacks and physically realizable attacks simultaneously. In this paper, we propose one of the first defense solutions against multiple adversarial video types for video classification. The proposed approach performs adversarial training with multiple types of video adversaries using independent batch normalizations (BNs), and recognizes different adversaries by an adversarial video detector. During inference, a switch module sends an input to a proper batch normalization branch according to the detected attack type. Compared to conventional adversarial training, our method exhibits stronger robustness to multiple and even unforeseen adversarial videos and provides higher classification accuracy.

Inspired by the strong correlation between the Label Smoothing Regularization(LSR) and Knowledge distillation(KD), we propose an algorithm LsrKD for training boost by extending the LSR method to the KD regime and applying a softer temperature. Then we improve the LsrKD by a Teacher Correction(TC) method, which manually sets a constant larger proportion for the right class in the uniform distribution teacher. To further improve the performance of LsrKD, we develop a self-distillation method named Memory-replay Knowledge Distillation (MrKD) that provides a knowledgeable teacher to replace the uniform distribution one in LsrKD. The MrKD method penalizes the KD loss between the current model's output distributions and its copies' on the training trajectory. By preventing the model learning so far from its historical output distribution space, MrKD can stabilize the learning and find a more robust minimum. Our experiments show that LsrKD can improve LSR performance consistently at no cost, especially on several deep neural networks where LSR is ineffectual. Also, MrKD can significantly improve single model training. The experiment results confirm that the TC can help LsrKD and MrKD to boost training, especially on the networks they are failed. Overall, LsrKD, MrKD, and their TC variants are comparable to or outperform the LSR method, suggesting the broad applicability of these KD methods.

Current deep learning research is dominated by benchmark evaluation. A method is regarded as favorable if it empirically performs well on the dedicated test set. This mentality is seamlessly reflected in the resurfacing area of continual learning, where consecutively arriving sets of benchmark data are investigated. The core challenge is framed as protecting previously acquired representations from being catastrophically forgotten due to the iterative parameter updates. However, comparison of individual methods is nevertheless treated in isolation from real world application and typically judged by monitoring accumulated test set performance. The closed world assumption remains predominant. It is assumed that during deployment a model is guaranteed to encounter data that stems from the same distribution as used for training. This poses a massive challenge as neural networks are well known to provide overconfident false predictions on unknown instances and break down in the face of corrupted data. In this work we argue that notable lessons from open set recognition, the identification of statistically deviating data outside of the observed dataset, and the adjacent field of active learning, where data is incrementally queried such that the expected performance gain is maximized, are frequently overlooked in the deep learning era. Based on these forgotten lessons, we propose a consolidated view to bridge continual learning, active learning and open set recognition in deep neural networks. Our results show that this not only benefits each individual paradigm, but highlights the natural synergies in a common framework. We empirically demonstrate improvements when alleviating catastrophic forgetting, querying data in active learning, selecting task orders, while exhibiting robust open world application where previously proposed methods fail.

Privacy issues were raised in the process of training deep learning in medical, mobility, and other fields. To solve this problem, we present privacy-preserving distributed deep learning method that allow clients to learn a variety of data without direct exposure. We divided a single deep learning architecture into a common extractor, a cloud model and a local classifier for the distributed learning. First, the common extractor, which is used by local clients, extracts secure features from the input data. The secure features also take the role that the cloud model can employ various task and diverse types of data. The feature contain the most important information that helps to proceed various task. Second, the cloud model including most parts of the whole training model gets the embedded features from the massive local clients, and performs most of deep learning operations which takes severe computing cost. After the operations in cloud model finished, outputs of the cloud model send back to local clients. Finally, the local classifier determined classification results and delivers the results to local clients. When clients train models, our model does not directly expose sensitive information to exterior network. During the test, the average performance improvement was 2.63% over the existing local training model. However, in a distributed environment, there is a possibility of inversion attack due to exposed features. For this reason, we experimented with the common extractor to prevent data restoration. The quality of restoration of the original image was tested by adjusting the depth of the common extractor. As a result, we found that the deeper the common extractor, the restoration score decreased to 89.74.

Multivariate time series (MTS) forecasting is an important problem in many fields. Accurate forecasting results can effectively help decision-making. Variables in MTS have rich relations among each other and the value of each variable in MTS depends both on its historical values and on other variables. These rich relations can be static and predictable or dynamic and latent. Existing methods do not incorporate these rich relational information into modeling or only model certain relation among MTS variables. To jointly model rich relations among variables and temporal dependencies within the time series, a novel end-to-end deep learning model, termed Multivariate Time Series Forecasting via Heterogenous Graph Neural Networks (MTHetGNN) is proposed in this paper. To characterize rich relations among variables, a relation embedding module is introduced in our model, where each variable is regarded as a graph node and each type of edge represents a specific relationship among variables or one specific dynamic update strategy to model the latent dependency among variables. In addition, convolutional neural network (CNN) filters with different perception scales are used for time series feature extraction, which is used to generate the feature of each node. Finally, heterogenous graph neural networks are adopted to handle the complex structural information generated by temporal embedding module and relation embedding module. Three benchmark datasets from the real world are used to evaluate the proposed MTHetGNN and the comprehensive experiments show that MTHetGNN achieves state-of-the-art results in MTS forecasting task.

Graph similarity computation, which predicts a similarity score between one pair of graphs, has been widely used in various fields, such as recommendation system Wu, Xiao and Chen (2015); Hu, Xu, Wang, Li and Liu (2020), computer vision Horaud and Skordas (1989); Pelillo, Siddiqi and Zucker (1999) and so on. However, most of the common distance measures evaluating how similar two graphs are, like Graph Edit Distance (GED) Bunke (1983) and Maximum Common Subgraph (MCS) Bunke and Shearer (1998), still suffer from large search spaces or excessive memory requirements. They are weak to compute exact graph distance for graphs with more than 16 nodes Blumenthal and Gamper (2018). Traditional graph similarity computation methods such as A* Riesen, Emmenegger and Bunke (2013), Hungarian Kuhn (1955); Riesen and Bunke (2009), VJ Fankhauser, Riesen and Bunke (2011); Jonker and Volgenant (1987), and BeamNeuhaus, Riesen and Bunke (2006), try to use pruning strategy or find approximate values instead of exact similarity to alleviate the problem. Nevertheless, by performing directly from the edges and nodes characteristics of the graphs, these exact and approximate algorithms still have a high time-complexity for computing the GED or MCS between two graphs, and are hard to be generalized to large graphs in real applications. With the rapid development of deep learning technology, graph embedding that automatically extracts the structural characteristics of the graph, provides a new solution for similarity computation and matching of graph structures. Recently, researchers proposed some representative graph deep learning models based on graph embedding for graph similarity computation.

The ability of intelligent agents to learn and remember multiple tasks sequentially is crucial to achieving artificial general intelligence. Many continual learning (CL) methods have been proposed to overcome catastrophic forgetting which results from non i.i.d data in the sequential learning of neural networks. In this paper we focus on class incremental learning, a challenging CL scenario. For this scenario, generative replay is a promising strategy which generates and replays pseudo data for previous tasks to alleviate catastrophic forgetting. However, it is hard to train a generative model continually for relatively complex data. Based on recently proposed orthogonal weight modification (OWM) algorithm which can approximately keep previously learned feature invariant when learning new tasks, we propose to 1) replay penultimate layer feature with a generative model; 2) leverage a self-supervised auxiliary task to further enhance the stability of feature. Empirical results on several datasets show our method always achieves substantial improvement over powerful OWM while conventional generative replay always results in a negative effect. Meanwhile our method beats several strong baselines including one based on real data storage. In addition, we conduct experiments to study why our method is effective.

Relation Detection is a task to determine whether two entities are related or not. In this paper, we employ neural network to do relation detection between two named entities for Indonesian Language. We used feature such as word embedding, position embedding, POS-Tag embedding, and character embedding. For the model, we divide the model into two parts: Front-part classifier (Convolutional layer or LSTM layer) and Back-part classifier (Dense layer or SVM). We did grid search method of neural network hyper parameter and SVM. We used 6000 Indonesian sentences for training process and 1,125 for testing. The best result is 0.8083 on F1-Score using Convolutional Layer as front-part and SVM as back-part.