The AAAI 2023 greatest paper awards have been introduced on the convention on Saturday 11 February. The awards comprised one excellent paper, one excellent scholar paper, and 12 distinguished papers.
AAAI-23 excellent paper
The AAAI excellent paper award is given to a paper (or papers) that “exemplifies the very best requirements in technical contribution and exposition”. This 12 months, the award goes to:
Misspecification in Inverse Reinforcement LearningJoar Skalse, Alessandro Abate
Summary: The goal of Inverse Reinforcement Studying (IRL) is to deduce a reward operate R from a coverage pi. To do that, we want a mannequin of how pi pertains to R. Within the present literature, the most typical fashions are optimality, Boltzmann rationality, and causal entropy maximisation. One of many major motivations behind IRL is to deduce human preferences from human behaviour. Nonetheless, the true relationship between human preferences and human behaviour is rather more advanced than any of the fashions presently utilized in IRL. Because of this they’re misspecified, which raises the concern that they could result in unsound inferences if utilized to real-world knowledge. On this paper, we offer a mathematical evaluation of how strong totally different IRL fashions are to misspecification, and reply exactly how the demonstrator coverage might differ from every of the usual fashions earlier than that mannequin results in defective inferences concerning the reward operate R. We additionally introduce a framework for reasoning about misspecification in IRL, along with formal instruments that can be utilized to simply derive the misspecification robustness of recent IRL fashions.
Learn the complete paper on arXiv.
AAAI-23 excellent scholar paper
An award to recognise an excellent contribution from a scholar. The 2023 winner is:
Beautify the Newcomers: Visible Area Immediate for Continuous Take a look at Time Adaptation Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, Lin Luo
Summary: Continuous Take a look at-Time Adaptation (CTTA) goals to adapt the supply mannequin to repeatedly altering unlabeled goal domains with out entry to the supply knowledge. Present strategies primarily deal with model-based adaptation in a self-training method, comparable to predicting pseudo labels for brand spanking new area datasets. Since pseudo labels are noisy and unreliable, these strategies endure from catastrophic forgetting and error accumulation when coping with dynamic knowledge distributions. Motivated by the immediate studying in NLP, on this paper, we suggest to study an image-layer visible area immediate for goal domains whereas having the supply mannequin parameters frozen. Throughout testing, the altering goal datasets might be tailored to the supply mannequin by reformulating the enter knowledge with the realized visible prompts. Particularly, we devise two sorts of prompts, i.e., domains-specific prompts and domains-agnostic prompts, to extract present area data and preserve the domain-shared data within the continuous adaptation. Moreover, we design a homeostasis-based adaptation technique to suppress domain-sensitive parameters in domain-invariant prompts to study domain-shared data extra successfully. This transition from the model-dependent paradigm to the model-free one permits us to bypass the catastrophic forgetting and error accumulation issues. Experiments present that our proposed technique achieves important efficiency beneficial properties over state-of-the-art strategies on 4 widely-used benchmarks, together with CIFAR-10C, CIFAR-100C, ImageNet-C, and VLCS datasets.
Learn the complete paper on arXiv.
AAAI-23 distinguished papers
The distinguished paper awards spotlight work which has been chosen for particular recognition. There are 12 winners this 12 months:
DropMessage: Unifying Random Dropping for Graph Neural NetworksTaoran Fang, Zhiqing Xiao, Chunping Wang, Jiarong Xu, Xuan Yang, Yang Yang
Summary: Graph Neural Networks (GNNs) are highly effective instruments for graph illustration studying. Regardless of their fast improvement, GNNs additionally face some challenges, comparable to over-fitting, over-smoothing, and non-robustness. Earlier works point out that these issues might be alleviated by random dropping strategies, which combine augmented knowledge into fashions by randomly masking components of the enter. Nonetheless, some open issues of random dropping on GNNs stay to be solved. First, it’s difficult to discover a common technique which might be appropriate for all circumstances contemplating the divergence of various datasets and fashions. Second, augmented knowledge launched to GNNs causes the unfinished protection of parameters and unstable coaching course of. Third, there is no such thing as a theoretical evaluation on the effectiveness of random dropping strategies on GNNs. On this paper, we suggest a novel random dropping technique referred to as DropMessage, which performs dropping operations instantly on the propagated messages in the course of the message-passing course of. Extra importantly, we discover that DropMessage gives a unified framework for many current random dropping strategies, based mostly on which we give theoretical evaluation of their effectiveness. Moreover, we elaborate the prevalence of DropMessage: it stabilizes the coaching course of by lowering pattern variance; it retains data range from the attitude of data principle, enabling it turn into a theoretical higher certain of different strategies. To judge our proposed technique, we conduct experiments that goals for a number of duties on 5 public datasets and two industrial datasets with numerous spine fashions. The experimental outcomes present that DropMessage has the benefits of each effectiveness and generalization, and might considerably alleviate the issues talked about above. An in depth model with full appendix might be discovered on arXiv.
Learn the complete paper on arXiv.
Two Heads are Higher than One: Picture-Level Cloud Community for Depth-Primarily based 3D Hand Pose EstimationPengfei Ren, Yuchen Chen, Jiachang Hao, Haifeng Solar, Qi Qi, Jingyu Wang, Jianxin Liao
Summary: Depth pictures and level clouds are the 2 mostly used knowledge representations for depth-based 3D hand pose estimation. Benefiting from the structuring of picture knowledge and the inherent inductive biases of the 2D Convolutional Neural Community (CNN), image-based strategies are extremely environment friendly and efficient. Nonetheless, treating the depth knowledge as a 2D picture inevitably ignores the 3D nature of depth knowledge. Level cloud-based strategies can higher mine the 3D geometric construction of depth knowledge. Nonetheless, these strategies endure from the dysfunction and non-structure of level cloud knowledge, which is computationally inefficient. On this paper, we suggest an Picture-Level cloud Community (IPNet) for correct and strong 3D hand pose estimation. IPNet makes use of 2D CNN to extract visible representations in 2D picture house and performs iterative correction in 3D level cloud house to use the 3D geometry data of depth knowledge. Specifically, we suggest a sparse anchor-based “aggregation-interaction-propagation” paradigm to reinforce level cloud options and refine the hand pose, which reduces irregular knowledge entry. Moreover, we introduce a 3D hand mannequin to the iterative correction course of, which considerably improves the robustness of IPNet to occlusion and depth holes. Experiments present that IPNet outperforms state-of-the-art strategies on three difficult hand datasets.
Neural Structure Seek for Huge Spectrum Adversarial RobustnessZhi Cheng, Yanxi Li, Minjing Dong, Xiu Su, Shan You, Chang Xu
Summary: One main limitation of CNNs is that they’re susceptible to adversarial assaults. At present, adversarial robustness in neural networks is often optimized with respect to a small pre-selected adversarial noise power, inflicting them to have doubtlessly restricted efficiency when underneath assault by bigger adversarial noises in real-world situations. On this analysis, we goal to search out Neural Architectures which have improved robustness on a variety of adversarial noise strengths by way of Neural Structure Search. Intimately, we suggest a light-weight Adversarial Noise Estimator to cut back the excessive price of producing adversarial noise with respect to totally different strengths. In addition to, we assemble an Environment friendly Huge Spectrum Searcher to cut back the price of adjusting community structure with the big adversarial validation set in the course of the search. With the 2 elements proposed, the variety of adversarial noise strengths searched might be elevated considerably whereas having a restricted improve in search time. In depth experiments on benchmark datasets comparable to CIFAR and ImageNet show that with a considerably richer search sign in robustness, our technique can discover architectures with improved total robustness whereas having a restricted impression on pure accuracy and round 40% discount in search time in contrast with the naive method of looking. Codes obtainable right here.
CowClip: Decreasing CTR Prediction Mannequin Coaching Time from 12 hours to 10 minutes on 1 GPUZangwei Zheng, Pengtai Xu, Xuan Zou, Da Tang, Zhen Li, Chenguang Xi, Peng Wu, Leqi Zou, Yijie Zhu, Ming Chen, Xiangzhuo Ding, Fuzhao Xue, Ziheng Qin, Youlong Cheng, Yang You
Summary: The clicking-through price (CTR) prediction process is to foretell whether or not a consumer will click on on the advisable merchandise. As mind-boggling quantities of information are produced on-line day by day, accelerating CTR prediction mannequin coaching is important to making sure an up-to-date mannequin and lowering the coaching price. One method to extend the coaching pace is to use giant batch coaching. Nonetheless, as proven in pc imaginative and prescient and pure language processing duties, coaching with a big batch simply suffers from the lack of accuracy. Our experiments present that earlier scaling guidelines fail within the coaching of CTR prediction neural networks. To deal with this downside, we first theoretically present that totally different frequencies of ids make it difficult to scale hyperparameters when scaling the batch measurement. To stabilize the coaching course of in a big batch measurement setting, we develop the adaptive Column-wise Clipping (CowClip). It permits a straightforward and efficient scaling rule for the embeddings, which retains the training price unchanged and scales the L2 loss. We conduct in depth experiments with 4 CTR prediction networks on two real-world datasets and efficiently scaled 128 instances the unique batch measurement with out accuracy loss. Specifically, for CTR prediction mannequin DeepFM coaching on the Criteo dataset, our optimization framework enlarges the batch measurement from 1K to 128K with over 0.1% AUC enchancment and reduces coaching time from 12 hours to 10 minutes on a single V100 GPU. Our code locates right here.
Learn the complete paper on arXiv.
DICNet: Deep Occasion-Degree Contrastive Community for Double Incomplete Multi-View MultiLabel ClassificationChengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, Yong Xu
Summary: In recent times, multi-view multi-label studying has aroused in depth analysis enthusiasm. Nonetheless, multi-view multi-label knowledge in the actual world is often incomplete as a result of unsure elements of information assortment and guide annotation, which implies that not solely multi-view options are sometimes lacking, and label completeness can also be tough to be happy. To take care of the double incomplete multi-view multi-label classification downside, we suggest a deep instance-level contrastive community, specifically DICNet. Completely different from standard strategies, our DICNet focuses on leveraging deep neural community to use the high-level semantic representations of samples fairly than shallow-level options. First, we make the most of the stacked autoencoders to construct an end-to-end multi-view function extraction framework to study the view-specific representations of samples. Moreover, to be able to enhance the consensus illustration capability, we introduce an incomplete instance-level contrastive studying scheme to information the encoders to higher extract the consensus data of a number of views and use a multi-view weighted fusion module to reinforce the discrimination of semantic options. General, our DICNet is adept in capturing constant discriminative representations of multi-view multi-label knowledge and avoiding the unfavorable results of lacking views and lacking labels. In depth experiments carried out on 5 datasets validate that our technique outperforms different state-of-the-art strategies.
Exploring Tuning Traits of Ventral Stream’s Neurons for Few-Shot ImageClassificationLintao Dong, Wei Zhai Zheng-Jun Zha
Summary: Human has the outstanding capability of studying novel objects by searching extraordinarily few examples, which can be attributed to the generic and strong function extracted within the ventral stream of our mind for representing visible objects. On this sense, the tuning traits of ventral stream’s neurons might be helpful prior data to enhance few-shot classification. Particularly, we computationally mannequin two teams of neurons present in ventral stream that are respectively delicate to form cues and colour cues. Then we suggest the hierarchical function regularization technique with these neuron fashions to regularize the spine of a few-shot mannequin, thus making it producing extra generic and strong options for few-shot classification. As well as, to simulate the tuning attribute that neuron firing at a better price in response to foreground stimulus components in comparison with background components, which we name belongingness, we design a foreground segmentation algorithm based mostly on the remark that the foreground object often doesn’t seem on the fringe of the image, then multiply the foreground masks with the spine of few-shot mannequin. Our technique is model-agnostic and might be utilized to few-shot fashions with totally different backbones, coaching paradigms and classifiers.
MaskBooster: Finish-to-Finish Self-Coaching for Sparsely Supervised Occasion SegmentationShida Zheng, Chenshu Chen, Xi Yang, Wenming Tan
Summary: The current paper introduces sparsely supervised occasion segmentation, with the datasets being totally annotated bounding bins and sparsely annotated masks. A direct answer to this process is self-training, which isn’t totally explored as an illustration segmentation but. On this paper, we suggest MaskBooster for sparsely supervised occasion segmentation (SpSIS) with complete utilization of pseudo masks. MaskBooster is featured with (1) dynamic and progressive pseudo masks from a web based updating instructor mannequin, (2) refining binary pseudo masks with the assistance of bounding field prior, (3) studying inter-class prediction distribution through data distillation for tender pseudo masks. As an end-to-end and common self-training framework, MaskBooster can empower totally supervised algorithms and increase their segmentation efficiency on SpSIS. Considerable experiments are performed on COCO and BDD100K datasets and validate the effectiveness of MaskBooster. Particularly, on COCO 0.1%/1%/10% protocols and BDD100K, we surpass sparsely supervised baseline by a big margin for each Masks RCNN and ShapeProp. MaskBooster on SpSIS additionally outperforms weakly and semi-supervised occasion segmentation state-of-the-art on the datasets with comparable annotation budgets.
SimFair: A Unified Framework for Equity-Conscious Multi-Label ClassificationTianci Liu, Haoyu Wang, Yaqing Wang, Xiaoqian Wang, Lu Su, Jing Gao
Summary: Current years have witnessed growing issues in direction of unfair selections made by machine studying algorithms. To enhance equity in mannequin selections, numerous equity notions have been proposed and lots of fairness-aware strategies are developed. Nonetheless, most of current definitions and strategies focus solely on single-label classification. Equity for multi-label classification, the place every occasion is related to multiple labels, remains to be but to determine. To fill this hole, we examine fairness-aware multi-label classification on this paper. We begin by extending Demographic Parity (DP) and Equalized Alternative (EOp), two common equity notions, to multi-label classification situations. Via a scientific examine, we present that on multi-label knowledge, due to inconsistently distributed labels, EOp often fails to assemble a dependable estimate on labels with few cases. We then suggest a brand new framework named Similarity s-induced Equity (sγ -SimFair). This new framework makes use of knowledge which have comparable labels when estimating equity on a specific label group for higher stability, and might unify DP and EOp. Theoretical evaluation and experimental outcomes on real-world datasets collectively show the benefit of sγ -SimFair over current strategies on multi-label classification duties.
XRand: Differentially Personal Protection in opposition to Rationalization-Guided AttacksTruc Nguyen, Phung Lai, Hai Phan, My T. Thai
Summary: Current improvement within the discipline of explainable synthetic intelligence (XAI) has helped enhance belief in Machine-Studying-as-a-Service (MLaaS) techniques, during which an evidence is supplied along with the mannequin prediction in response to every question. Nonetheless, XAI additionally opens a door for adversaries to achieve insights into the black-box fashions in MLaaS, thereby making the fashions extra susceptible to a number of assaults. For instance, feature-based explanations (e.g., SHAP) might expose the highest vital options {that a} black-box mannequin focuses on. Such disclosure has been exploited to craft efficient backdoor triggers in opposition to malware classifiers. To deal with this trade-off, we introduce a brand new idea of reaching native differential privateness (LDP) within the explanations, and from that we set up a protection, referred to as XRand, in opposition to such assaults. We present that our mechanism restricts the data that the adversary can study concerning the prime vital options, whereas sustaining the faithfulness of the reasons.
Learn the complete paper on arXiv.
Clustering What Issues: Optimum Approximation for Clustering with OutliersAkanksha Agrawal, Tanmay Inamdar, Saket Saurabh, Jie Xue
Summary: Clustering with outliers is without doubt one of the most basic issues in Laptop Science. Given a set 
of 
factors and two numbers 
, the clustering with outliers goals to exclude 
factors from and partition the remaining factors into 
clusters that minimizes a sure price operate. On this paper, we give a basic method for fixing clustering with outliers, which leads to a fixed-parameter tractable (FPT) algorithm in and , that just about matches the approximation ratio for its outlier-free counterpart. As a corollary, we acquire FPT approximation algorithms with optimum approximation ratios for -MEDIAN and -MEANS with outliers typically and Euclidean metrics. We additionally exhibit extra functions of our method to different variants of the issue that impose extra constraints on the clustering, comparable to equity or matroid constraints.
Learn the complete paper on arXiv.
Strong Common-Reward Markov Choice ProcessesYue Wang, Alvaro Velasquez, George Atia, Ashley Prater-Bennette, Shaofeng Zou
Summary: In strong Markov resolution processes (MDPs), the uncertainty within the transition kernel is addressed by discovering a coverage that optimizes the worst-case efficiency over an uncertainty set of MDPs. Whereas a lot of the literature has targeted on discounted MDPs, strong average-reward MDPs stay largely unexplored. On this paper, we deal with strong average-reward MDPs, the place the aim is to discover a coverage that optimizes the worst-case common reward over an uncertainty set. We first take an method that approximates average-reward MDPs utilizing discounted MDPs. We show that the strong discounted worth operate converges to the strong average-reward because the low cost issue 
goes to 
, and furthermore, when is giant, any optimum coverage of the strong discounted MDP can also be an optimum coverage of the strong average-reward. We additional design a strong dynamic programming method, and theoretically characterize its convergence to the optimum. Then, we examine strong average-reward MDPs instantly with out utilizing discounted MDPs as an intermediate step. We derive the strong Bellman equation for strong average-reward MDPs, show that the optimum coverage might be derived from its answer, and additional design a strong relative worth iteration algorithm that provably discover its answer, or equivalently, the optimum strong coverage.
Learn the complete paper on arXiv.
Environment friendly Reply Enumeration in Description Logics with Useful RolesCarsten Lutz, Marcin Przybylko
Summary: We examine the enumeration of solutions to ontology-mediated queries when the ontology is formulated in an outline logic that helps useful roles and the question is a CQ. Specifically, we present that enumeration is feasible with linear preprocessing and fixed delay when a sure extension of the CQ (pertaining to useful roles) is acyclic and free-connex acyclic. This holds each for full solutions and for partial solutions. We offer matching decrease bounds for the case the place the question is self-join free.
Learn an prolonged model of this paper on arXiv.
tags: AAAI2023
Lucy Smith
, Managing Editor for AIhub.
