We propose a novel approach for visual representation learning called Signature-Graph Neural Networks (SGN). SGN learns latent global structures that augment the feature representation of Convolutional Neural Networks (CNN). SGN constructs unique undirected graphs for each image based on the CNN feature maps. The feature maps are partitioned into a set of equal and non-overlapping patches. The graph nodes are located on high-contrast sharp convolution features with the local maxima or minima in these patches. The node embeddings are aggregated through novel Signature-Graphs based on horizontal and vertical edge connections. The representation vectors are then computed based on the spectral Laplacian eigenvalues of the graphs. SGN outperforms existing methods of recent graph convolutional networks, generative adversarial networks, and auto-encoders with image classification accuracy of 99.65% on ASIRRA, 99.91% on MNIST, 98.55% on Fashion-MNIST, 96.18% on CIFAR-10, 84.71% on CIFAR-100, 94.36% on STL10, and 95.86% on SVHN datasets. We also introduce a novel implementation of the state-of-the-art multi-head attention (MHA) on top of the proposed SGN. Adding SGN to MHA improved the image classification accuracy from 86.92% to 94.36% on the STL10 dataset
Recently, tremendous human-designed and automatically searched neural networks have been applied to image denoising. However, previous works intend to handle all noisy images in a pre-defined static network architecture, which inevitably leads to high computational complexity for good denoising quality. Here, we present dynamic slimmable denoising network (DDS-Net), a general method to achieve good denoising quality with less computational complexity, via dynamically adjusting the channel configurations of networks at test time with respect to different noisy images. Our DDS-Net is empowered with the ability of dynamic inference by a dynamic gate, which can predictively adjust the channel configuration of networks with negligible extra computation cost. To ensure the performance of each candidate sub-network and the fairness of the dynamic gate, we propose a three-stage optimization scheme. In the first stage, we train a weight-shared slimmable super network. In the second stage, we evaluate the trained slimmable super network in an iterative way and progressively tailor the channel numbers of each layer with minimal denoising quality drop. By a single pass, we can obtain several sub-networks with good performance under different channel configurations. In the last stage, we identify easy and hard samples in an online way and train a dynamic gate to predictively select the corresponding sub-network with respect to different noisy images. Extensive experiments demonstrate our DDS-Net consistently outperforms the state-of-the-art individually trained static denoising networks.
Convolutional neural networks (CNNs) have been successfully applied to the single target tracking task in recent years. Generally, training a deep CNN model requires numerous labeled training samples, and the number and quality of these samples directly affect the representational capability of the trained model. However, this approach is restrictive in practice, because manually labeling such a large number of training samples is time-consuming and prohibitively expensive. In this paper, we propose an active learning method for deep visual tracking, which selects and annotates the unlabeled samples to train the deep CNNs model. Under the guidance of active learning, the tracker based on the trained deep CNNs model can achieve competitive tracking performance while reducing the labeling cost. More specifically, to ensure the diversity of selected samples, we propose an active learning method based on multi-frame collaboration to select those training samples that should be and need to be annotated. Meanwhile, considering the representativeness of these selected samples, we adopt a nearest neighbor discrimination method based on the average nearest neighbor distance to screen isolated samples and low-quality samples. Therefore, the training samples subset selected based on our method requires only a given budget to maintain the diversity and representativeness of the entire sample set. Furthermore, we adopt a Tversky loss to improve the bounding box estimation of our tracker, which can ensure that the tracker achieves more accurate target states. Extensive experimental results confirm that our active learning-based tracker (ALT) achieves competitive tracking accuracy and speed compared with state-of-the-art trackers on the seven most challenging evaluation benchmarks.
Multimedia event detection is the task of detecting a specific event of interest in an user-generated video on websites. The most fundamental challenge facing this task lies in the enormously varying quality of the video as well as the high-level semantic abstraction of event inherently. In this paper, we decompose the video into several segments and intuitively model the task of complex event detection as a multiple instance learning problem by representing each video as a "bag" of segments in which each segment is referred to as an instance. Instead of treating the instances equally, we associate each instance with a reliability variable to indicate its importance and then select reliable instances for training. To measure the reliability of the varying instances precisely, we propose a visual-semantic guided loss by exploiting low-level feature from visual information together with instance-event similarity based high-level semantic feature. Motivated by curriculum learning, we introduce a negative elastic-net regularization term to start training the classifier with instances of high reliability and gradually taking the instances with relatively low reliability into consideration. An alternative optimization algorithm is developed to solve the proposed challenging non-convex non-smooth problem. Experimental results on standard datasets, i.e., TRECVID MEDTest 2013 and TRECVID MEDTest 2014, demonstrate the effectiveness and superiority of the proposed method to the baseline algorithms.
Dynamic networks have shown their promising capability in reducing theoretical computation complexity by adapting their architectures to the input during inference. However, their practical runtime usually lags behind the theoretical acceleration due to inefficient sparsity. Here, we explore a hardware-efficient dynamic inference regime, named dynamic weight slicing, which adaptively slice a part of network parameters for inputs with diverse difficulty levels, while keeping parameters stored statically and contiguously in hardware to prevent the extra burden of sparse computation. Based on this scheme, we present dynamic slimmable network (DS-Net) and dynamic slice-able network (DS-Net++) by input-dependently adjusting filter numbers of CNNs and multiple dimensions in both CNNs and transformers, respectively. To ensure sub-network generality and routing fairness, we propose a disentangled two-stage optimization scheme with training techniques such as in-place bootstrapping (IB), multi-view consistency (MvCo) and sandwich gate sparsification (SGS) to train supernet and gate separately. Extensive experiments on 4 datasets and 3 different network architectures demonstrate our method consistently outperforms state-of-the-art static and dynamic model compression methods by a large margin (up to 6.6%). Typically, DS-Net++ achieves 2-4x computation reduction and 1.62x real-world acceleration over MobileNet, ResNet-50 and Vision Transformer, with minimal accuracy drops (0.1-0.3%) on ImageNet. Code release: https://github.com/changlin31/DS-Net
Political stance detection has become an important task due to the increasingly polarized political ideologies. Most existing works focus on identifying perspectives in news articles or social media posts, while social entities, such as individuals and organizations, produce these texts and actually take stances. In this paper, we propose the novel task of entity stance prediction, which aims to predict entities' stances given their social and political context. Specifically, we retrieve facts from Wikipedia about social entities regarding contemporary U.S. politics. We then annotate social entities' stances towards political ideologies with the help of domain experts. After defining the task of entity stance prediction, we propose a graph-based solution, which constructs a heterogeneous information network from collected facts and adopts gated relational graph convolutional networks for representation learning. Our model is then trained with a combination of supervised, self-supervised and unsupervised loss functions, which are motivated by multiple social and political phenomenons. We conduct extensive experiments to compare our method with existing text and graph analysis baselines. Our model achieves highest stance detection accuracy and yields inspiring insights regarding social entity stances. We further conduct ablation study and parameter analysis to study the mechanism and effectiveness of our proposed approach.
Identifying political perspective in news media has become an important task due to the rapid growth of political commentary and the increasingly polarized ideologies. Previous approaches only focus on leveraging the semantic information and leaves out the rich social and political context that helps individuals understand political stances. In this paper, we propose a perspective detection method that incorporates external knowledge of real-world politics. Specifically, we construct a contemporary political knowledge graph with 1,071 entities and 10,703 triples. We then build a heterogeneous information network for each news document that jointly models article semantics and external knowledge in knowledge graphs. Finally, we apply gated relational graph convolutional networks and conduct political perspective detection as graph-level classification. Extensive experiments show that our method achieves the best performance and outperforms state-of-the-art methods by 5.49%. Numerous ablation studies further bear out the necessity of external knowledge and the effectiveness of our graph-based approach.
Zero-shot object detection (ZSD), the task that extends conventional detection models to detecting objects from unseen categories, has emerged as a new challenge in computer vision. Most existing approaches tackle the ZSD task with a strict mapping-transfer strategy, which may lead to suboptimal ZSD results: 1) the learning process of those models ignores the available unseen class information, and thus can be easily biased towards the seen categories; 2) the original visual feature space is not well-structured and lack of discriminative information. To address these issues, we develop a novel Semantics-Guided Contrastive Network for ZSD, named ContrastZSD, a detection framework that first brings contrastive learning mechanism into the realm of zero-shot detection. Particularly, ContrastZSD incorporates two semantics-guided contrastive learning subnets that contrast between region-category and region-region pairs respectively. The pairwise contrastive tasks take advantage of additional supervision signals derived from both ground truth label and pre-defined class similarity distribution. Under the guidance of those explicit semantic supervision, the model can learn more knowledge about unseen categories to avoid the bias problem to seen concepts, while optimizing the data structure of visual features to be more discriminative for better visual-semantic alignment. Extensive experiments are conducted on two popular benchmarks for ZSD, i.e., PASCAL VOC and MS COCO. Results show that our method outperforms the previous state-of-the-art on both ZSD and generalized ZSD tasks.
Navigation is one of the fundamental features of a autonomous robot. And the ability of long-term navigation with semantic instruction is a `holy grail` goals of intelligent robots. The development of 3D simulation technology provide a large scale of data to simulate the real-world environment. The deep learning proves its ability to robustly learn various embodied navigation tasks. However, deep learning on embodied navigation is still in its infancy due to the unique challenges faced by the navigation exploration and learning from partial observed visual input. Recently, deep learning in embodied navigation has become even thriving, with numerous methods have been proposed to tackle different challenges in this area. To give a promising direction for future research, in this paper, we present a comprehensive review of embodied navigation tasks and the recent progress in deep learning based methods. It includes two major tasks: target-oriented navigation and the instruction-oriented navigation.
With leveraging the weight-sharing and continuous relaxation to enable gradient-descent to alternately optimize the supernet weights and the architecture parameters through a bi-level optimization paradigm, \textit{Differentiable ARchiTecture Search} (DARTS) has become the mainstream method in Neural Architecture Search (NAS) due to its simplicity and efficiency. However, more recent works found that the performance of the searched architecture barely increases with the optimization proceeding in DARTS. In addition, several concurrent works show that the NAS could find more competitive architectures without labels. The above observations reveal that the supervision signal in DARTS may be a poor indicator for architecture optimization, inspiring a foundational question: instead of using the supervision signal to perform bi-level optimization, \textit{can we find high-quality architectures \textbf{without any training nor labels}}? We provide an affirmative answer by customizing the NAS as a network pruning at initialization problem. By leveraging recent techniques on the network pruning at initialization, we designed a FreeFlow proxy to score the importance of candidate operations in NAS without any training nor labels, and proposed a novel framework called \textit{training and label free neural architecture search} (\textbf{FreeNAS}) accordingly. We show that, without any training nor labels, FreeNAS with the proposed FreeFlow proxy can outperform most NAS baselines. More importantly, our framework is extremely efficient, which completes the architecture search within only \textbf{3.6s} and \textbf{79s} on a single GPU for the NAS-Bench-201 and DARTS search space, respectively. We hope our work inspires more attempts in solving NAS from the perspective of pruning at initialization.