IJCAI 2021 Papers

The main challenge of cross-modal retrieval is to learn the consistent embedding for heterogeneous modalities. To solve this problem, traditional label-wise cross-modal approaches usually constrain the inter-modal and intra-modal embedding consistency relying on the label ground-truths. However, the experiments reveal that different modal networks actually have various generalization capacities, thereby end-to-end joint training with consistency loss usually leads to sub-optimal uni-modal model, which in turn affects the learning of consistent embedding. Therefore, in this paper, we argue that what really needed for supervised cross-modal retrieval is a good shared classification model. In other words, we learn the consistent embedding by ensuring the classification performance of each modality on the shared model, without the consistency loss. Specifically, we consider a technique called Semantic Sharing, which directly trains the two modalities interactively by adopting a shared self-attention based classification model. We evaluate the proposed approach on three representative datasets. The results validate that the proposed semantic sharing can consistently boost the performance under NDCG metric.

Over-parameterized deep neural networks (DNNs) with sufficient capacity to memorize random noise can achieve excellent generalization performance, challenging the bias-variance trade-off in classical learning theory. Recent studies claimed that DNNs first learn simple patterns and then memorize noise; some other works showed a phenomenon that DNNs have a spectral bias to learn target functions from low to high frequencies during training. However, we show that the monotonicity of the learning bias does not always hold: under the experimental setup of deep double descent, the high-frequency components of DNNs diminish in the late stage of training, leading to the second descent of the test error. Besides, we find that the spectrum of DNNs can be applied to indicating the second descent of the test error, even though it is calculated from the training set only.

Retrosynthesis, of which the goal is to find a set of reactants for synthesizing a target product, is an emerging research area of deep learning. While the existing approaches have shown promising results, they currently lack the ability to consider availability (e.g., stability or purchasability) of the reactants or generalize to unseen reaction templates (i.e., chemical reaction rules). In this paper, we propose a new approach that mitigates the issues by reformulating retrosynthesis into a selection problem of reactants from a candidate set of commercially available molecules. To this end, we design an efficient reactant selection framework, named RetCL (retrosynthesis via contrastive learning), for enumerating all of the candidate molecules based on selection scores computed by graph neural networks. For learning the score functions, we also propose a novel contrastive training scheme with hard negative mining. Extensive experiments demonstrate the benefits of the proposed selection-based approach. For example, when all 671k reactants in the USPTO database are given as candidates, our RetCL achieves top-1 exact match accuracy of 71.3% for the USPTO-50k benchmark, while a recent transformer-based approach achieves 59.6%. We also demonstrate that RetCL generalizes well to unseen templates in various settings in contrast to template-based approaches.

An accurate short-term electric load forecasting is critical for modern electric power systems' safe and economical operation. Electric load forecasting can be formulated as a multi-variate time series problem. Residential houses in the same neighborhood may be affected by similar factors and share some latent spatial dependencies. However, most of the existing works on electric load forecasting fail to explore such dependencies. In recent years, graph neural networks (GNNs) have shown impressive success in modeling such dependencies. However, such GNN based models usually would require a large amount of training data. We may have a minimal amount of data available to train a reliable forecasting model for houses in a new neighborhood area. At the same time, we may have a large amount of historical data collected from other houses that can be leveraged to improve the new neighborhood's prediction performance. In this paper, we propose an attentive transfer learning-based GNN model that can utilize the learned prior knowledge to improve the learning process in a new area. The transfer process is achieved by an attention network, which generically avoids negative transfer by leveraging knowledge from multiple sources. Extensive experiments have been conducted on real-world data sets. Results have shown that the proposed framework can consistently outperform baseline models in different areas.

Deep hashing methods have shown great retrieval accuracy and efficiency in large-scale image retrieval. How to optimize discrete hash bits is always the focus in deep hashing methods. A common strategy in these methods is to adopt an activation function, e.g. sigmoid() or tanh(), and minimize a quantization loss to approximate discrete values. However, this paradigm may make more and more hash bits stuck into the wrong saturated area of the activation functions and never escaped. We call this problem "Dead Bits Problem (DBP)". Besides, the existing quantization loss will aggravate DBP as well. In this paper, we propose a simple but effective gradient amplifier which acts before activation functions to alleviate DBP. Moreover, we devise an error-aware quantization loss to further alleviate DBP. It avoids the negative effect of quantization loss based on the similarity between two images. The proposed gradient amplifier and error-aware quantization loss are compatible with a variety of deep hashing methods. Experimental results on three datasets demonstrate the efficiency of the proposed gradient amplifier and the error-aware quantization loss.

Foreground occlusion removal task aims to automatically detect and remove foreground occlusions and recover background objects. Since for Light Fields (LFs), background objects occluded in some views may be seen in other views, the foreground occlusion removal task for LFs is easy to achieve. In this paper, we propose a learning-based method combining ‘seeking’ and ‘generating’ to recover occluded background. Specifically, the micro-lens dynamic filters are proposed to ‘seek’ occluded background points in shifted micro-lens images and remove occlusions using angular information. The shifted images are then combined to further ‘generate’ background regions to supplement more background details using spatial information. By fully exploring the angular and spatial information in LFs, the dense and complex occlusions can be easily removed. Quantitative and qualitative experimental results show that our method outperforms other state-of-the-arts methods by a large margin.

The core is a well-known and fundamental notion of stability in games intended to model coalition formation such as hedonic games. The fact that the number of deviating agents (that have to coordinate themselves) can be arbitrarily high, and the fact that agents may benefit only by a tiny amount from their deviation (while they could incur in a cost for deviating), suggest that the core is not able to suitably model many practical scenarios in large and highly distributed multi-agent systems. For this reason, we consider relaxed core stable outcomes where the notion of permissible deviations is modified along two orthogonal directions: the former takes into account the size of the deviating coalition, and the latter the amount of utility gain for each member of the deviating coalition. These changes result in two different notions of stability, namely, the q-size core and k-improvement core. We investigate these concepts of stability in fractional hedonic games, that is a well-known subclass of hedonic games for which core stable outcomes are not guaranteed to exist and it is computationally hard to decide nonemptiness of the core. Interestingly, the considered relaxed notions of core also possess the appealing property of recovering, in some notable cases, the convergence, the existence and the possibility of computing stable solutions in polynomial time.

This paper addresses the challenge of learning to do procedural reasoning over text to answer "What if..." questions. We propose a novel relational gating network that learns to filter the key entities and relationships and learns contextual and cross representations of both procedure and question for finding the answer. Our relational gating network contains an entity gating module, relation gating module, and contextual interaction module. These modules help in solving the "What if..." reasoning problem. We show that modeling pairwise relationships helps to capture higher-order relations and find the line of reasoning for causes and effects in the procedural descriptions. Our proposed approach achieves the state-of-the-art results on the WIQA dataset.

Learning how to execute complex tasks involving multiple objects in a 3D world is challenging when there is no ground-truth information about the objects or any demonstration to learn from. When an agent only receives a signal from task-completion, this makes it challenging to learn the object-representations which support learning the correct object-interactions needed to complete the task. In this work, we formulate learning an attentive object dynamics model as a classification problem, using random object-images to define incorrect labels for our object-dynamics model. We show empirically that this enables object-representation learning that captures an object's category (is it a toaster?), its properties (is it on?), and object-relations (is something inside of it?). With this, our core learner (a relational RL agent) receives the dense training signal it needs to rapidly learn object-interaction tasks. We demonstrate results in the 3D AI2Thor simulated kitchen environment with a range of challenging food preparation tasks. We compare our method's performance to several related approaches and against the performance of an oracle: an agent that is supplied with ground-truth information about objects in the scene. We find that our agent achieves performance closest to the oracle in terms of both learning speed and maximum success rate.

Application of deep learning to NP-hard combinatorial optimization problems is an emerging research trend, and a number of interesting approaches have been published over the last few years. In this work we address robust optimization, which is a more complex variant where a max-min problem is to be solved. We obtain robust solutions by solving the inner minimization problem exactly and apply Reinforcement Learning to learn a heuristic for the outer problem. The minimization term in the inner objective represents an obstacle to existing RL-based approaches, as its value depends on the full solution in a non-linear manner and cannot be evaluated for partial solutions constructed by the agent over the course of each episode. We overcome this obstacle by defining the reward in terms of the one-step advantage over a baseline policy whose role can be played by any fast heuristic for the given problem. The agent is trained to maximize the total advantage, which, as we show, is equivalent to the original objective. We validate our approach by solving min-max versions of standard benchmarks for the Capacitated Vehicle Routing and the Traveling Salesperson Problem, where our agents obtain near-optimal solutions and improve upon the baselines.

We explore the black-box adversarial attack on video recognition models. Attacks are only performed on selected key regions and key frames to reduce the high computation cost of searching adversarial perturbations on a video due to its high dimensionality. To select key frames, one way is to use heuristic algorithms to evaluate the importance of each frame and choose the essential ones. However, it is time inefficient on sorting and searching. In order to speed up the attack process, we propose a reinforcement learning based frame selection strategy. Specifically, the agent explores the difference between the original class and the target class of videos to make selection decisions. It receives rewards from threat models which indicate the quality of the decisions. Besides, we also use saliency detection to select key regions and only estimate the sign of gradient instead of the gradient itself in zeroth order optimization to further boost the attack process. We can use the trained model directly in the untargeted attack or with little fine-tune in the targeted attack, which saves computation time. A range of empirical results on real datasets demonstrate the effectiveness and efficiency of the proposed method.

As one of the most popular generative models, Variational Autoencoder (VAE) approximates the posterior of latent variables based on amortized variational inference. However, when the decoder network is sufficiently expressive, VAE may lead to posterior collapse; that is, uninformative latent representations may be learned. To this end, in this paper, we propose an alternative model, DU-VAE, for learning a more Diverse and less Uncertain latent space, and thus the representation can be learned in a meaningful and compact manner. Specifically, we first theoretically demonstrate that it will result in better latent space with high diversity and low uncertainty awareness by controlling the distribution of posterior’s parameters across the whole data accordingly. Then, without the introduction of new loss terms or modifying training strategies, we propose to exploit Dropout on the variances and Batch-Normalization on the means simultaneously to regularize their distributions implicitly. Furthermore, to evaluate the generalization effect, we also exploit DU-VAE for inverse autoregressive flow based-VAE (VAE-IAF) empirically. Finally, extensive experiments on three benchmark datasets clearly show that our approach can outperform state-of-the-art baselines on both likelihood estimation and underlying classification tasks.

Machine learning classifiers’ capability is largely dependent on the scale of available training data and limited by the model overfitting in data-scarce learning tasks. To address this problem, this work proposes a novel Meta Functional Learning (MFL) by meta-learning a generalisable functional model from data-rich tasks whilst simultaneously regularising knowledge transfer to data-scarce tasks. The MFL computes meta-knowledge on functional regularisation generalisable to different learning tasks by which functional training on limited labelled data promotes more discriminative functions to be learned. Moreover, we adopt an Iterative Update strategy on MFL (MFL-IU). This improves knowledge transfer regularisation from MFL by progressively learning the functional regularisation in knowledge transfer. Experiments on three Few-Shot Learning (FSL) benchmarks (miniImageNet, CIFAR-FS and CUB) show that meta functional learning for regularisation knowledge transfer can benefit improving FSL classifiers.

In the literature, we find reductions from 3SAT to Max2SAT. These reductions are based on the usage of a gadget, i.e., a combinatorial structure that allows translating constraints of one problem to constraints of another. Unfortunately, the generation of these gadgets lacks an intuitive or efficient method. In this paper, we provide an efficient and constructive method for Reducing SAT to Max2SAT and show empirical results of how MaxSAT solvers are more efficient than SAT solvers solving the translation of hard formulas for Resolution.

The bus system is a critical component of sustainable urban transportation. However, due to the significant uncertainties in passenger demand and traffic conditions, bus operation is unstable in nature and bus bunching has become a common phenomenon that undermines the reliability and efficiency of bus services. Despite recent advances in multi-agent reinforcement learning (MARL) on traffic control, little research has focused on bus fleet control due to the tricky asynchronous characteristic---control actions only happen when a bus arrives at a bus stop and thus agents do not act simultaneously. In this study, we formulate route-level bus fleet control as an asynchronous multi-agent reinforcement learning (ASMR) problem and extend the classical actor-critic architecture to handle the asynchronous issue. Specifically, we design a novel critic network to effectively approximate the marginal contribution for other agents, in which graph attention neural network is used to conduct inductive learning for policy evaluation. The critic structure also helps the ego agent optimize its policy more efficiently. We evaluate the proposed framework on real-world bus services and actual passenger demand derived from smart card data. Our results show that the proposed model outperforms both traditional headway-based control methods and existing MARL methods.

Predictive state representations (PSRs) are models of controlled non-Markov observation sequences which exhibit the same generative process governing POMDP observations without relying on an underlying latent state. In that respect, a PSR is indistinguishable from the corresponding POMDP. However, PSRs notoriously ignore the notion of rewards, which undermines the general utility of PSR models for control, planning, or reinforcement learning. Therefore, we describe a sufficient and necessary accuracy condition which determines whether a PSR is able to accurately model POMDP rewards, we show that rewards can be approximated even when the accuracy condition is not satisfied, and we find that a non-trivial number of POMDPs taken from a well-known third-party repository do not satisfy the accuracy condition. We propose reward-predictive state representations (R-PSRs), a generalization of PSRs which accurately models both observations and rewards, and develop value iteration for R-PSRs. We show that there is a mismatch between optimal POMDP policies and the optimal PSR policies derived from approximate rewards. On the other hand, optimal R-PSR policies perfectly match optimal POMDP policies, reconfirming R-PSRs as accurate state-less generative models of observations and rewards.

Word Sense Disambiguation (WSD) aims at making explicit the semantics of a word in context by identifying the most suitable meaning from a predefined sense inventory. Recent breakthroughs in representation learning have fueled intensive WSD research, resulting in considerable performance improvements, breaching the 80% glass ceiling set by the inter-annotator agreement. In this survey, we provide an extensive overview of current advances in WSD, describing the state of the art in terms of i) resources for the task, i.e., sense inventories and reference datasets for training and testing, as well as ii) automatic disambiguation approaches, detailing their peculiarities, strengths and weaknesses. Finally, we highlight the current limitations of the task itself, but also point out recent trends that could help expand the scope and applicability of WSD, setting up new promising directions for the future.

Recommender systems have played a critical role in many web applications to meet user's personalized interests and alleviate the information overload. In this survey, we review the development of recommendation frameworks with the focus on heterogeneous relational learning, which consists of different types of dependencies among users and items. The objective of this task is to map heterogeneous relational data into latent representation space, such that the structural and relational properties from both user and item domain can be well preserved. To address this problem, recent research developments can fall into three major categories: social recommendation, knowledge graph-enhanced recommender system, and multi-behavior recommendation. We discuss the learning approaches in each category, such as matrix factorization, attention mechanism and graph neural networks, for effectively distilling heterogeneous contextual information. Finally, we present exploratory outlook to highlight several promising directions and opportunities in heterogeneous relational learning frameworks for recommendation.

Adversarial training is one of the most effective approaches for deep learning models to defend against adversarial examples. Unlike other defense strategies, adversarial training aims to enhance the robustness of models intrinsically. During the past few years, adversarial training has been studied and discussed from various aspects, which deserves a comprehensive review. For the first time in this survey, we systematically review the recent progress on adversarial training for adversarial robustness with a novel taxonomy. Then we discuss the generalization problems in adversarial training from three perspectives and highlight the challenges which are not fully tackled. Finally, we present potential future directions.