IJCAI 2022 Papers

Most real-world optimization problems have multiple objectives. A system designer needs to find a policy that trades off these objectives to reach a desired operating point. This problem has been studied extensively in the setting of known objective functions. However, we consider a more practical but challenging setting of unknown objective functions. In industry, optimization under this setting is mostly approached with online A/B testing, which is often costly and inefficient. As an alternative, we propose Interactive Multi-Objective Off-policy Optimization (IMO^3). The key idea of IMO^3 is to interact with a system designer using policies evaluated in an off-policy fashion to uncover which policy maximizes her unknown utility function. We theoretically show that IMO^3 identifies a near-optimal policy with high probability, depending on the amount of designer's feedback and training data for off-policy estimation. We demonstrate its effectiveness empirically on several multi-objective optimization problems.

Face Super-resolution (FSR) task works for generating high-resolution (HR) face images from the corresponding low-resolution (LR) inputs, which has received a lot of attentions because of the wide application prospects. However, due to the diversity of facial texture and the difficulty of reconstructing detailed content from degraded images, FSR technology is still far away from being solved. In this paper, we propose a novel and effective face super-resolution framework based on Transformer, namely Interconnected Dual Pyramid Transformer (IDPT). Instead of straightly stacking cascaded feature reconstruction blocks, the proposed IDPT designs the pyramid encoder/decoder Transformer architecture to extract coarse and detailed facial textures respectively, while the relationship between the dual pyramid Transformers is further explored by a bottom pyramid feature extractor. The pyramid encoder/decoder structure is devised to adapt various characteristics of textures in different spatial spaces hierarchically. A novel fusing modulation module is inserted in each spatial layer to guide the refinement of detailed texture by the corresponding coarse texture, while fusing the shallow-layer coarse feature and corresponding deep-layer detailed feature simultaneously. Extensive experiments and visualizations on various datasets demonstrate the superiority of the proposed method for face super-resolution tasks.

Precise segmentation of polyps from colonoscopic images is extremely significant for the early diagnosis and treatment of colorectal cancer. However, it is still a challenging task due to: (1)the boundary between the polyp and the background is blurred makes delineation difficult; (2)the various size and shapes causes feature representation of polyps difficult. In this paper, we propose an integration context-based reverse-contour guidance network (ICGNet) to solve these challenges. The ICGNet firstly utilizes a reverse-contour guidance module to aggregate low-level edge detail information and meanwhile constraint reverse region. Then, the newly designed adaptive context module is used to adaptively extract local-global information of the current layer and complementary information of the previous layer to get larger and denser features. Lastly, an innovative hybrid pyramid pooling fusion module fuses the multi-level features generated from the decoder in the case of considering salient features and less background. Our proposed approach is evaluated on the EndoScene, Kvasir-SEG and CVC-ColonDB datasets with eight evaluation metrics, and gives competitive results compared with other state-of-the-art methods in both learning ability and generalization capability.

The quality of a blastocyst directly determines the embryo's implantation potential, thus making it essential to objectively and accurately identify the blastocyst morphology. In this work, we propose an automatic framework named I2CNet to perform the blastocyst segmentation task in human embryo images. The I2CNet contains two components: IntrA-Class Context Module (IACCM) and InteR-Class Context Module (IRCCM). The IACCM aggregates the representations of specific areas sharing the same category for each pixel, where the categorized regions are learned under the supervision of the groundtruth. This aggregation decomposes a K-category recognition task into K recognition tasks of two labels while maintaining the ability of garnering intra-class features. In addition, the IRCCM is designed based on the blastocyst morphology to compensate for inter-class information which is gradually gathered from inside out. Meanwhile, a weighted mapping function is applied to facilitate edges of the inter classes and stimulate some hard samples. Eventually, the learned intra- and inter-class cues are integrated from coarse to fine, rendering sufficient information interaction and fusion between multi-scale features. Quantitative and qualitative experiments demonstrate that the superiority of our model compared with other representative methods. The I2CNet achieves accuracy of 94.14% and Jaccard of 85.25% on blastocyst public dataset.

We study mechanisms that allocate reviewers to papers in a fair and efficient manner. We model reviewer assignment as an instance of a fair allocation problem, presenting an extension of the classic round-robin mechanism, called Reviewer Round Robin (RRR). Round-robin mechanisms are a standard tool to ensure envy-free up to one item (EF1) allocations. However, fairness often comes at the cost of decreased efficiency. To overcome this challenge, we carefully select an approximately optimal round-robin order. Applying a relaxation of submodularity, γ-weak submodularity, we show that greedily inserting papers into an order yields a (1+γ²)-approximation to the maximum welfare attainable by our round-robin mechanism under any order. Our Greedy Reviewer Round Robin (GRRR) approach outputs highly efficient EF1 allocations for three real conference datasets, offering comparable performance to state-of-the-art paper assignment methods in fairness, efficiency, and runtime, while providing the only EF1 guarantee.

Forecasting the future trajectories of multiple agents is a core technology for human-robot interaction systems. To predict multi-agent trajectories more accurately, it is inevitable that models need to improve interpretability and reduce redundancy. However, many methods adopt implicit weight calculation or black-box networks to learn the semantic interaction of agents, which obviously lack enough interpretation. In addition, most of the existing works model the relation among all agents in a one-to-one manner, which might lead to irrational trajectory predictions due to its redundancy and noise. To address the above issues, we present Hypertron, a human-understandable and lightweight hypergraph-based multi-agent forecasting framework, to explicitly estimate the motions of multiple agents and generate reasonable trajectories. The framework explicitly interacts among multiple agents and learns their latent intentions by our coarse-to-fine hypergraph convolution interaction module. Our experiments on several challenging real-world trajectory forecasting datasets show that Hypertron outperforms a wide array of state-of-the-art methods while saving over 60% parameters and reducing 30% inference time.

Hypergraphs are natural and expressive modeling tools to encode high-order relationships among entities. Several variations of Hypergraph Neural Networks (HGNNs) are proposed to learn the node representations and complex relationships in the hypergraphs. Most current approaches assume that the input hypergraph structure accurately depicts the relations in the hypergraphs. However, the input hypergraph structure inevitably contains noise, task-irrelevant information, or false-negative connections. Treating the input hypergraph structure as ground-truth information unavoidably leads to sub-optimal performance. In this paper, we propose a Hypergraph Structure Learning (HSL) framework, which optimizes the hypergraph structure and the HGNNs simultaneously in an end-to-end way. HSL learns an informative and concise hypergraph structure that is optimized for downstream tasks. To efficiently learn the hypergraph structure, HSL adopts a two-stage sampling process: hyperedge sampling for pruning redundant hyperedges and incident node sampling for pruning irrelevant incident nodes and discovering potential implicit connections. The consistency between the optimized structure and the original structure is maintained by the intra-hyperedge contrastive learning module. The sampling processes are jointly optimized with HGNNs towards the objective of the downstream tasks. Experiments conducted on 7 datasets show shat HSL outperforms the state-of-the-art baselines while adaptively sparsifying hypergraph structures.

Few-shot learning (FSL) aims to recognize a novel class with very few instances, which is a challenging task since it suffers from a data scarcity issue. One way to effectively alleviate this issue is introducing explicit knowledge summarized from human past experiences to achieve knowledge transfer for FSL. Based on this idea, in this paper, we introduce the explicit knowledge of class hierarchy (i.e., the hierarchy relations between classes) as FSL priors and propose a novel hyperbolic knowledge transfer framework for FSL, namely, HyperKT. Our insight is, in the hyperbolic space, the hierarchy relation between classes can be well preserved by resorting to the exponential growth characters of hyperbolic volume, so that better knowledge transfer can be achieved for FSL. Specifically, we first regard the class hierarchy as a tree-like structure. Then, 1) a hyperbolic representation learning module and a hyperbolic prototype inference module are employed to encode/infer each image and class prototype to the hyperbolic space, respectively; and 2) a novel hierarchical classification and relation reconstruction loss are carefully designed to learn the class hierarchy. Finally, the novel class prediction is performed in a nearest-prototype manner. Extensive experiments on three datasets show our method achieves superior performance over state-of-the-art methods, especially on 1-shot tasks.

Adequate annotated data cannot always be satisfied in medical imaging applications. To address such a challenge, we would explore ways to reduce the quality and quantity of annotations requirements of the deep learning model by developing a hybrid learning system. We combined self-supervised learning, semi-supervised learning and weak-supervised learning to improve annotation utilization. Our primary research work on 2D medical image detection under poor annotation conditions has found that better regularization and adversarial loss can improve the robustness and performance with poor annotation conditions.

Most of today's AI systems focus on using self-attention mechanisms and transformer architectures on large amounts of diverse data to achieve impressive performance gains. In this paper, we propose to augment the transformer architecture with an external attention mechanism to bring external knowledge and context to bear. By integrating external information into the prediction process, we hope to reduce the need for ever-larger models and increase the democratization of AI systems. We find that the proposed external attention mechanism can significantly improve the performance of existing AI systems, allowing practitioners to easily customize foundation AI models to many diverse downstream applications. In particular, we focus on the task of Commonsense Reasoning, demonstrating that the proposed external attention mechanism can augment existing transformer models and significantly improve the model's reasoning capabilities. The proposed system, Knowledgeable External Attention for commonsense Reasoning (KEAR), reaches human parity on the open CommonsenseQA research benchmark with an accuracy of 89.4% in comparison to the human accuracy of 88.9%.

We extend the map-of-elections framework to the case of approval elections. While doing so, we study a number of statistical cultures, including some new ones, and we analyze their properties. We find that approval elections can be understood in terms of the average number of approvals in the votes, and the extent to which the votes are chaotic.

Voting is a crucial methodology for eliciting and combining agents' preferences and information across many applications. Just as there are numerous voting rules exhibiting different properties, we also see many different voting systems. In this paper we investigate how different voting systems perform as a function of the characteristics of the underlying voting population and social network. In particular, we compare direct democracy, liquid democracy, and sortition in a ground truth voting context. Through simulations -- using both real and artificially generated social networks -- we illustrate how voter competency distributions and levels of direct participation affect group accuracy differently in each voting mechanism. Our results can be used to guide the selection of a suitable voting system based on the characteristics of a particular voting setting.

Model robustness is vital for the reliable deployment of machine learning models in real-world applications. Recent studies have shown that data augmentation can result in model over-relying on features in the low-frequency domain, sacrificing performance against low-frequency corruptions, highlighting a connection between frequency and robustness. Here, we take one step further to more directly study the frequency bias of a model through the lens of its Jacobians and its implication to model robustness. To achieve this, we propose Jacobian frequency regularization for models' Jacobians to have a larger ratio of low-frequency components. Through experiments on four image datasets, we show that biasing classifiers towards low (high)-frequency components can bring performance gain against high (low)-frequency corruption and adversarial perturbation, albeit with a tradeoff in performance for low (high)-frequency corruption. Our approach elucidates a more direct connection between the frequency bias and robustness of deep learning models.

Expectation maximization (EM) is the default algorithm for fitting probabilistic models with missing or latent variables, yet we lack a full understanding of its non-asymptotic convergence properties. Previous works show results along the lines of “EM converges at least as fast as gradient descent” by assuming the conditions for the convergence of gradient descent apply. This approach is not only loose, in that it does not capture that EM can make more progress than a gradient step, but the assumptions fail to hold for textbook examples of EM like Gaussian mixtures. In this work, we show that for the common setting of exponential family distributions, viewing EM as a mirror descent algorithm leads to convergence rates in Kullback-Leibler (KL) divergence and how the KL divergence is related to first-order stationarity via Bregman divergences. In contrast to previous works, the analysis is invariant to the choice of parametrization and holds with minimal assumptions. We also show applications of these ideas to local linear (and superlinear) convergence rates, generalized EM, and non-exponential family distributions.

Multilingual neural machine translation (MNMT) trained in multiple language pairs has attracted considerable attention due to fewer model parameters and lower training costs by sharing knowledge among multiple languages. Nonetheless, multilingual training is plagued by language interference degeneration in shared parameters because of the negative interference among different translation directions, especially on high-resource languages. In this paper, we propose the multilingual translation model with the high-resource language-specific training (HLT-MT) to alleviate the negative interference, which adopts the two-stage training with the language-specific selection mechanism. Specifically, we first train the multilingual model only with the high-resource pairs and select the language-specific modules at the top of the decoder to enhance the translation quality of high-resource directions. Next, the model is further trained on all available corpora to transfer knowledge from high-resource languages (HRLs) to low-resource languages (LRLs). Experimental results show that HLT-MT outperforms various strong baselines on WMT-10 and OPUS-100 benchmarks. Furthermore, the analytic experiments validate the effectiveness of our method in mitigating the negative interference in multilingual training.

This paper outlines an automated creation system for Chinese landscape paintings based on generative adversarial networks. The system consists of three cascaded modules: generation, resizing, and super-resolution. The generation module first generates a square-shaped painting, then the resizing module predicts an appropriate aspect ratio for it and performs resizing, and finally the super-resolution module is used to increase the resolution and improve the quality. After training each module with the images we collected from the web, our system can create high-resolution landscape paintings in arbitrary sizes.

We propose HifiHead, a high fidelity neural talking head synthesis method, which can well preserve the source image's appearance and control the motion (e.g., pose, expression, gaze) flexibly with 3D morphable face models (3DMMs) parameters derived from a driving image or indicated by users. Existing head synthesis works mainly focus on low-resolution inputs. Instead, we exploit the powerful generative prior embedded in StyleGAN to achieve high-quality head synthesis and editing. Specifically, we first extract the source image's appearance and driving image's motion to construct 3D face descriptors, which are employed as latent style codes for the generator. Meanwhile, hierarchical representations are extracted from the source and rendered 3D images respectively to provide faithful appearance and shape guidance. Considering the appearance representations need high-resolution flow fields for spatial transform, we propose a coarse-to-fine style-based generator consisting of a series of feature alignment and refinement (FAR) blocks. Each FAR block updates the dense flow fields and refines RGB outputs simultaneously for efficiency. Extensive experiments show that our method blends source appearance and target motion more accurately along with more photo-realistic results than previous state-of-the-art approaches.

Graph neural network (GNN) is popular now to solve the tasks in non-Euclidean space and most of them learn deep embeddings by aggregating the neighboring nodes. However, these methods are prone to some problems such as over-smoothing because of the single-scale perspective field and the nature of low-pass filter. To address these limitations, we introduce diffusion scattering network (DSN) to exploit high-order patterns. With observing the complementary relationship between multi-layer GNN and DSN, we propose Hierarchical Diffusion Scattering Graph Neural Network (HDS-GNN) to efficiently bridge DSN and GNN layer by layer to supplement GNN with multi-scale information and band-pass signals. Our model extracts node-level scattering representations by intercepting the low-pass filtering, and adaptively tunes the different scales to regularize multi-scale information. Then we apply hierarchical representation enhancement to improve GNN with the scattering features. We benchmark our model on nine real-world networks on the transductive semi-supervised node classification task. The experimental results demonstrate the effectiveness of our method.

In the past few decades, Hadamard-based image restoration problems (e.g., low-light image enhancement) attract wide concerns in multiple areas related to artificial intelligence. However, existing works mostly focus on heuristically defining architecture and loss by the engineering experiences that came from extensive practices. This way brings about expensive verification costs for seeking out the optimal solution. To this end, we develop a novel hierarchical bilevel learning scheme to discover the architecture and loss simultaneously for different Hadamard-based image restoration tasks. More concretely, we first establish a new Hadamard-inspired neural unit to aggregate domain knowledge into the network design. Then we model a triple-level optimization that consists of the architecture, loss and parameters optimizations to deliver a macro perspective for network learning. Then we introduce a new hierarchical bilevel learning scheme for solving the built triple-level model to progressively generate the desired architecture and loss. We also define an architecture search space consisting of a series of simple operations and an image quality-oriented loss search space. Extensive experiments on three Hadamard-based image restoration tasks (including low-light image enhancement, single image haze removal and underwater image enhancement) fully verify our superiority against state-of-the-art methods.