AAAI 2022 Papers

Transfer adversarial attack is a non-trivial black-box adversarial attack that aims to craft adversarial perturbations on the surrogate model and then apply such perturbations to the victim model. However, the transferability of perturbations from existing methods is still limited, since the adversarial perturbations are easily overfitting with a single surrogate model and specific data pattern. In this paper, we propose a Learning to Learn Transferable Attack (LLTA) method, which makes the adversarial perturbations more generalized via learning from both data and model augmentation. For data augmentation, we adopt simple random resizing and padding. For model augmentation, we randomly alter the back propagation instead of the forward propagation to eliminate the effect on the model prediction. By treating the attack of both specific data and a modified model as a task, we expect the adversarial perturbations to adopt enough tasks for generalization. To this end, the meta-learning algorithm is further introduced during the iteration of perturbation generation. Empirical results on the widely-used dataset demonstrate the effectiveness of our attack method with a 12.85% higher success rate of transfer attack compared with the state-of-the-art methods. We also evaluate our method on the real-world online system, i.e., Google Cloud Vision API, to further show the practical potentials of our method.

The Elo rating system is widely adopted to evaluate the skills of (chess) game and sports players. Recently it has been also integrated into machine learning algorithms in evaluating the performance of computerised AI agents. However, an accurate estimation of the Elo rating (for the top players) often requires many rounds of competitions, which can be expensive to carry out. In this paper, to minimize the number of comparisons and to improve the sample efficiency of the Elo evaluation (for top players), we propose an efficient online match scheduling algorithm. Specifically, we identify and match the top players through a dueling bandits framework and tailor the bandit algorithm to the gradient-based update of Elo. We show that it reduces the per-step memory and time complexity to constant, compared to the traditional likelihood maximization approaches requiring O(t) time. Our algorithm has a regret guarantee that is sublinear in the number of competition rounds and has been extended to the multidimensional Elo ratings for handling intransitive games. We empirically demonstrate that our method achieves superior convergence speed and time efficiency on a variety of gaming tasks.

3D facial landmark detection is extensively used in many research fields such as face registration, facial shape analysis, and face recognition. Most existing methods involve traditional features and 3D face models for the detection of landmarks, and their performances are limited by the hand-crafted intermediate process. In this paper, we propose a novel 3D facial landmark detection method, which directly locates the coordinates of landmarks from 3D point cloud with a well-customized graph convolutional network. The graph convolutional network learns geometric features adaptively for 3D facial landmark detection with the assistance of constructed 3D heatmaps, which are Gaussian functions of distances to each landmark on a 3D face. On this basis, we further develop a local surface unfolding and registration module to predict 3D landmarks from the heatmaps. The proposed method forms the first baseline of deep point cloud learning method for 3D facial landmark detection. We demonstrate experimentally that the proposed method exceeds the existing approaches by a clear margin on BU-3DFE and FRGC datasets for landmark localization accuracy and stability, and also achieves high-precision results on a recent large-scale dataset.

We propose a new problem setting to study the sequential interactions between a recommender system and a user. Instead of assuming the user is omniscient, static, and explicit, as the classical practice does, we sketch a more realistic user behavior model, under which the user: 1) rejects recommendations if they are clearly worse than others; 2) updates her utility estimation based on rewards from her accepted recommendations; 3) withholds realized rewards from the system. We formulate the interactions between the system and such an explorative user in a K-armed bandit framework and study the problem of learning the optimal recommendation on the system side. We show that efficient system learning is still possible but is more difficult. In particular, the system can identify the best arm with probability at least 1-delta within O(1/delta) interactions, and we prove this is tight. Our finding contrasts the result for the problem of best arm identification with fixed confidence, in which the best arm can be identified with probability 1-delta within O(log(1/delta)) interactions. This gap illustrates the inevitable cost the system has to pay when it learns from an explorative user's revealed preferences on its recommendations rather than from the realized rewards.

Reasoning is a dynamic process. In cognitive theories, the dynamics of reasoning refers to reasoning states over time after successive state transitions. Modeling the cognitive dynamics is of utmost importance to simulate human reasoning capability. In this paper, we propose to learn the reasoning dynamics of visual relational reasoning by casting it as a path routing task. We present a reinforced path routing method that represents an input image via a structured visual graph and introduces a reinforcement learning based model to explore paths (sequences of nodes) over the graph based on an input sentence to infer reasoning results. By exploring such paths, the proposed method represents reasoning states clearly and characterizes state transitions explicitly to fully model the reasoning dynamics for accurate and transparent visual relational reasoning. Extensive experiments on referring expression comprehension and visual question answering demonstrate the effectiveness of our method.

Due to the scarcity of annotated samples, the diversity between support set and query set becomes the main obstacle for few shot semantic segmentation. Most existing prototype-based approaches only exploit the prototype from the support feature and ignore the information from the query sample, failing to remove this obstacle.In this paper, we proposes a dual prototype network (DPNet) to dispose of few shot semantic segmentation from a new perspective. Along with the prototype extracted from the support set, we propose to build the pseudo-prototype based on foreground features in the query image. To achieve this goal, the cycle comparison module is developed to select reliable foreground features and generate the pseudo-prototype with them. Then, a prototype interaction module is utilized to integrate the information of the prototype and the pseudo-prototype based on their underlying correlation. Finally, a multi-scale fusion module is introduced to capture contextual information during the dense comparison between prototype (pseudo-prototype) and query feature. Extensive experiments conducted on two benchmarks demonstrate that our method exceeds previous state-of-the-arts with a sizable margin, verifying the effectiveness of the proposed method.

Current pre-training methods in computer vision focus on natural images in the daily-life context. However, abstract diagrams such as icons and symbols are common and important in the real world. We are inspired by Tangram, a game that requires replicating an abstract pattern from seven dissected shapes. By recording human experience in solving tangram puzzles, we present the Tangram dataset and show that a pre-trained neural model on the Tangram helps solve some mini visual tasks based on low-resolution vision. Extensive experiments demonstrate that our proposed method generates intelligent solutions for aesthetic tasks such as folding clothes and evaluating room layouts. The pre-trained feature extractor can facilitate the convergence of few-shot learning tasks on human handwriting and improve the accuracy in identifying icons by their contours. The Tangram dataset is available at https://github.com/yizhouzhao/Tangram.

Graph Convolutional Neural Networks (GCNs) have been generally accepted to be an effective tool for node representations learning. An interesting way to understand GCNs is to think of them as a message passing mechanism where each node updates its representation by accepting information from its neighbours (also known as positive samples). However, beyond these neighbouring nodes, graphs have a large, dark, all-but forgotten world in which we find the non-neighbouring nodes (negative samples). In this paper, we show that this great dark world holds a substantial amount of information that might be useful for representation learning. Most specifically, it can provide negative information about the node representations. Our overall idea is to select appropriate negative samples for each node and incorporate the negative information contained in these samples into the representation updates. Moreover, we show that the process of selecting the negative samples is not trivial. Our theme therefore begins by describing the criteria for a good negative sample, followed by a determinantal point process algorithm for efficiently obtaining such samples. A GCN, boosted by diverse negative samples, then jointly considers the positive and negative information when passing messages. Experimental evaluations show that this idea not only improves the overall performance of standard representation learning but also significantly alleviates over-smoothing problems.

Learning visual knowledge from massive weakly-labeled web videos has attracted growing research interests thanks to the large corpus of easily accessible video data on the Internet. However, for video action recognition, the action of interest might only exist in arbitrary clips of untrimmed web videos, resulting in high label noises in the temporal space. To address this challenge, we introduce a new method for pre-training video action recognition models using queried web videos. Instead of trying to filter out potential noises, we propose to provide fine-grained supervision signals by defining the concept of Sub-Pseudo Label (SPL). Specifically, SPL spans out a new set of meaningful "middle ground" label space constructed by extrapolating the original weak labels during video querying and the prior knowledge distilled from a teacher model. Consequently, SPL provides enriched supervision for video models to learn better representations and improves data utilization efficiency of untrimmed videos. We validate the effectiveness of our method on four video action recognition datasets and a weakly-labeled image dataset. Experiments show that SPL outperforms several existing pre-training strategies and the learned representations lead to competitive results on several benchmarks.

Learning from one's mistakes is an effective human learning technique where the learners focus more on the topics where mistakes were made, so as to deepen their understanding. In this paper, we investigate if this human learning strategy can be applied in machine learning. We propose a novel machine learning method called Learning From Mistakes (LFM), wherein the learner improves its ability to learn by focusing more on the mistakes during revision. We formulate LFM as a three-stage optimization problem: 1) learner learns; 2) learner re-learns focusing on the mistakes, and; 3) learner validates its learning. We develop an efficient algorithm to solve the LFM problem. We apply the LFM framework to neural architecture search on CIFAR-10, CIFAR-100, and Imagenet. Experimental results strongly demonstrate the effectiveness of our model.

The use of priors to avoid manual labeling for training machine learning methods has received much attention in the last few years. One of the critical subthemes in this regard is Learning from Label Proportions (LLP), where only the information about class proportions is available for training the models. While various LLP training settings verse in the literature, most approaches focus on bag-level label proportions errors, often leading to suboptimal solutions. This paper proposes a new model that jointly uses prototypical contrastive learning and bag-level cluster proportions to implement efficient LLP classification. Our proposal explicitly relaxes the equipartition constraint commonly used in prototypical contrastive learning methods and incorporates the exact cluster proportions into the optimal transport algorithm used for cluster assignments. At inference time, we compute the clusters' assignment, delivering instance-level classification. We experimented with our method on two widely used image classification benchmarks and report a new state-of-art LLP performance, achieving results close to fully supervised methods.

An artificial neural network can be trained by uniformly broadcasting a reward signal to units that implement a REINFORCE learning rule. Though this presents a biologically plausible alternative to backpropagation in training a network, the high variance associated with it renders it impractical to train deep networks. The high variance arises from the inefficient structural credit assignment since a single reward signal is used to evaluate the collective action of all units. To facilitate structural credit assignment, we propose replacing the reward signal to hidden units with the change in the L2 norm of the unit's outgoing weight. As such, each hidden unit in the network is trying to maximize the norm of its outgoing weight instead of the global reward, and thus we call this learning method Weight Maximization. We prove that Weight Maximization is approximately following the gradient of rewards in expectation. In contrast to backpropagation, Weight Maximization can be used to train both continuous-valued and discrete-valued units. Moreover, Weight Maximization solves several major issues of backpropagation relating to biological plausibility. Our experiments show that a network trained with Weight Maximization can learn significantly faster than REINFORCE and slightly slower than backpropagation. Weight Maximization illustrates an example of cooperative behavior automatically arising from a population of self-interested agents in a competitive game without any central coordination.

Sub-seasonal forecasting (SSF) is the prediction of key climate variables such as temperature and precipitation on the 2-week to 2-month time horizon. Skillful SSF would have substantial societal value in areas such as agricultural productivity, hydrology and water resource management, and emergency planning for extreme events such as droughts and wildfires. Despite its societal importance, SSF has stayed a challenging problem compared to both short-term weather forecasting and long-term seasonal forecasting. Recent studies have shown the potential of machine learning (ML) models to advance SSF. In this paper, for the first time, we perform a fine-grained comparison of a suite of modern ML models with start-of-the-art physics-based dynamical models from the Subseasonal Experiment (SubX) project for SSF in the western contiguous United States. Additionally, we explore mechanisms to enhance the ML models by using forecasts from dynamical models. Empirical results illustrate that, on average, ML models outperform dynamical models while the ML models tend to generate forecasts with conservative magnitude compared to the SubX models. Further, we illustrate that ML models make forecasting errors under extreme weather conditions, e.g., cold waves due to the polar vortex, highlighting the need for separate models for extreme events. Finally, we show that suitably incorporating dynamical model forecasts as inputs to ML models can substantially improve the forecasting performance of the ML models. The SSF dataset constructed for the work and code for the ML models are released along with the paper for the benefit of the artificial intelligence community.

This paper proposes a representational model for image pairs such as consecutive video frames that are related by local pixel displacements, in the hope that the model may shed light on motion perception in primary visual cortex (V1). The model couples the following two components: (1) the vector representations of local contents of images and (2) the matrix representations of local pixel displacements caused by the relative motions between the agent and the objects in the 3D scene. When the image frame undergoes changes due to local pixel displacements, the vectors are multiplied by the matrices that represent the local displacements. Thus the vector representation is equivariant as it varies according to the local displacements. Our experiments show that our model can learn Gabor-like filter pairs of quadrature phases. The profiles of the learned filters match those of simple cells in Macaque V1. Moreover, we demonstrate that the model can learn to infer local motions in either a supervised or unsupervised manner. With such a simple model, we achieve competitive results on optical flow estimation.

We present a novel generalized zero-shot algorithm to recognize perceived emotions from gestures. Our task is to map gestures to novel emotion categories not encountered in training. We introduce an adversarial autoencoder-based representation learning that correlates 3D motion-captured gesture sequences with the vectorized representation of the natural-language perceived emotion terms using word2vec embeddings. The language-semantic embedding provides a representation of the emotion label space, and we leverage this underlying distribution to map the gesture sequences to the appropriate categorical emotion labels. We train our method using a combination of gestures annotated with known emotion terms and gestures not annotated with any emotions. We evaluate our method on the MPI Emotional Body Expressions Database (EBEDB) and obtain an accuracy of 58.43%. We see an improvement in performance compared to current state-of-the-art algorithms for generalized zero-shot learning by an absolute 25-27%. We also demonstrate our approach on publicly available online videos and movie scenes, where the actors' pose has been extracted and map to their respective emotive states.

Deep learning models have shown to be susceptible to universal adversarial perturbation (UAP), which has aroused wide concerns in the community. Compared with the conventional adversarial attacks that generate adversarial samples at the instance level, UAP can fool the target model for different instances with only a single perturbation, enabling us to evaluate the robustness of the model from a more effective and accurate perspective. The existing universal attack methods fail to exploit the differences and connections between the instance and universal levels to produce dominant perturbations. To address this challenge, we propose a new universal attack method that unifies instance-specific and universal attacks from a feature perspective to generate a more dominant UAP. Specifically, we reformulate the UAP generation task as a minimax optimization problem and then utilize the instance-specific attack method to solve the minimization problem thereby obtaining better training data for generating UAP. At the same time, we also introduce a consistency regularizer to explore the relationship between training data, thus further improving the dominance of the generated UAP. Furthermore, our method is generic with no additional assumptions about the training data and hence can be applied to both data-dependent (supervised) and data-independent (unsupervised) manners. Extensive experiments demonstrate that the proposed method improves the performance by a significant margin over the existing methods in both data-dependent and data-independent settings. Code is available at https://github.com/lisenxd/AT-UAP.

In image retrieval, deep local features learned in a data-driven manner have been demonstrated effective to improve retrieval performance. To realize efficient retrieval on large image database, some approaches quantize deep local features with a large codebook and match images with aggregated match kernel. However, the complexity of these approaches is non-trivial with large memory footprint, which limits their capability to jointly perform feature learning and aggregation. To generate compact global representations while maintaining regional matching capability, we propose a unified framework to jointly learn local feature representation and aggregation. In our framework, we first extract local features using CNNs. Then, we design a tokenizer module to aggregate them into a few visual tokens, each corresponding to a specific visual pattern. This helps to remove background noise, and capture more discriminative regions in the image. Next, a refinement block is introduced to enhance the visual tokens with self-attention and cross-attention. Finally, different visual tokens are concatenated to generate a compact global representation. The whole framework is trained end-to-end with image-level labels. Extensive experiments are conducted to evaluate our approach, which outperforms the state-of-the-art methods on the Revisited Oxford and Paris datasets.

Unpaired video-to-video translation aims to translate videos between a source and a target domain without the need of paired training data, making it more feasible for real applications. Unfortunately, the translated videos generally suffer from temporal and semantic inconsistency. To address this, many existing works adopt spatiotemporal consistency constraints incorporating temporal information based on motion estimation. However, the inaccuracies in the estimation of motion deteriorate the quality of the guidance towards spatiotemporal consistency, which leads to unstable translation. In this work, we propose a novel paradigm that regularizes the spatiotemporal consistency by synthesizing motions in input videos with the generated optical flow instead of estimating them. Therefore, the synthetic motion can be applied in the regularization paradigm to keep motions consistent across domains without the risk of errors in motion estimation. Thereafter, we utilize our unsupervised recycle and unsupervised spatial loss, guided by the pseudo-supervision provided by the synthetic optical flow, to accurately enforce spatiotemporal consistency in both domains. Experiments show that our method is versatile in various scenarios and achieves state-of-the-art performance in generating temporally and semantically consistent videos. Code is available at: https://github.com/wangkaihong/Unsup_Recycle_GAN/.

Recent developments in predictive modeling using marked temporal point processes (MTPPs) have enabled an accurate characterization of several real-world applications involving continuous-time event sequences (CTESs). However, the retrieval problem of such sequences remains largely unaddressed in literature. To tackle this, we propose NEUROSEQRET which learns to retrieve and rank a relevant set of continuous-time event sequences for a given query sequence, from a large corpus of sequences. More specifically, NEUROSEQRET first applies a trainable unwarping function on the query sequence, which makes it comparable with corpus sequences, especially when a relevant query-corpus pair has individually different attributes. Next, it feeds the unwarped query sequence and the corpus sequence into MTPP guided neural relevance models. We develop two variants of the relevance model which offer a tradeoff between accuracy and efficiency. We also propose an optimization framework to learn binary sequence embeddings from the relevance scores, suitable for the locality-sensitive hashing leading to a significant speedup in returning top-K results for a given query sequence. Our experiments with several datasets show the significant accuracy boost of NEUROSEQRET beyond several baselines, as well as the efficacy of our hashing mechanism.