AAAI 2024 论文列表

Large language models (LLMs) offer impressive performance in various zero-shot and few-shot tasks. However, their success in zero-shot or few-shot settings may be affected by task contamination, a potential limitation that has not been thoroughly examined. This paper investigates how zero-shot and few-shot performance of LLMs has changed chronologically over datasets released over time, and over LLMs released over time. Utilizing GPT-3 series models and several other recent open-sourced LLMs, and controlling for dataset difficulty, we find that datasets released prior to the LLM training data creation date perform surprisingly better than datasets released post the LLM training data creation date. This strongly indicates that, for many LLMs, there exists task contamination on zero-shot and few-shot evaluation for datasets prior to the LLMs' training data creation date. Additionally, we utilize training data inspection, training data extraction, and a membership inference attack, which reveal further evidence of task contamination. Importantly, we find that for tasks with no possibility of task contamination, LLMs rarely demonstrate statistically significant improvements over simple majority baselines, in both zero and few-shot settings.

Neural networks (NNs) can learn to rely on spurious signals in the training data, leading to poor generalisation. Recent methods tackle this problem by training NNs with additional ground-truth annotations of such signals. These methods may, however, let spurious signals re-emerge in deep convolutional NNs (CNNs). We propose Targeted Activation Penalty (TAP), a new method tackling the same problem by penalising activations to control the re-emergence of spurious signals in deep CNNs, while also lowering training times and memory usage. In addition, ground-truth annotations can be expensive to obtain. We show that TAP still works well with annotations generated by pre-trained models as effective substitutes of ground-truth annotations. We demonstrate the power of TAP against two state-of-the-art baselines on the MNIST benchmark and on two clinical image datasets, using four different CNN architectures.

The population characteristics of the datasets related to the same task may vary significantly and merging them may harm performance. In this paper, we propose a novel method of domain adaptation called "cross-adaptation". It allows for implicit adaptation to the target domain without the need for any labeled examples across this domain. We test our approach on 9 datasets for SARS-CoV-2 detection from complete blood count from different hospitals around the world. Results show that our solution is universal with respect to various classification algorithms and allows for up to a 10pp increase in F1 score on average.

Template learning transformer trackers have achieved significant performance improvement recently due to the longdependency learning using the self-attention (SA) mechanism. However, the typical SA mechanisms in transformers adopt a less discriminative design approach which is inadequate for focusing on the most important target information during tracking. Therefore, existing trackers are easily distracted by background information and have constraints in handling tracking challenges. The focus of our research is to develop a target-focused discriminative shallow transformer tracking framework that can learn to distinguish the target from the background and enable accurate tracking with fast speed. Extensive experiments will be performed on several popular benchmarks, including OTB100, UAV123, GOT10k, LaSOT, and TrackingNet, to demonstrate the effectiveness of the proposed framework.

Sigmoid output layers are widely used in multi-label classification (MLC) tasks, in which multiple labels can be assigned to any input. In many practical MLC tasks, the number of possible labels is in the thousands, often exceeding the number of input features and resulting in a low-rank output layer. In multi-class classification, it is known that such a low-rank output layer is a bottleneck that can result in unargmaxable classes: classes which cannot be predicted for any input. In this paper, we show that for MLC tasks, the analogous sigmoid bottleneck results in exponentially many unargmaxable label combinations. We explain how to detect these unargmaxable outputs and demonstrate their presence in three widely used MLC datasets. We then show that they can be prevented in practice by introducing a Discrete Fourier Transform (DFT) output layer, which guarantees that all sparse label combinations with up to k active labels are argmaxable. Our DFT layer trains faster and is more parameter efficient, matching the F1@k score of a sigmoid layer while using up to 50% fewer trainable parameters. Our code is publicly available at https://github.com/andreasgrv/sigmoid-bottleneck.

There has been a great deal of recent interest in binarized neural networks, especially because of their explainability. At the same time, automatic differentiation algorithms such as backpropagation fail for binarized neural networks, which limits their applicability. We show that binarized neural networks admit a tame representation by reformulating the problem of training binarized neural networks as a subadditive dual of a mixed-integer program, which we show to have nice properties. This makes it possible to use the framework of Bolte et al. for implicit differentiation, which offers the possibility for practical implementation of backpropagation in the context of binarized neural networks. This approach could also be used for a broader class of mixed-integer programs, beyond the training of binarized neural networks, as encountered in symbolic approaches to AI and beyond.

Humor is a crucial part of human communication. Understanding humor and generating humorous responses in dialogue can provide natural and empathic human-computer interactions. However, most existing pre-trained language models (PLMs) perform unsatisfactorily in humor generation. On one hand, the serious shortage of humor corpus and datasets pose challenges for constructing models that can understand and generate humorous expressions. On the other hand, humor generation relies on rich knowledge and commonsense, which is often tacit and unspoken. In this paper, we construct the largest Chinese Explainable Humor Response Dataset to date with chain-of-humor and humor mind map annotations, which can be used to comprehensively evaluate as well as improve the humorous response ability of PLMs. We further design humor-related auxiliary tasks to further enhance PLMs' humorous response performance. Extensive evaluations demonstrate that our proposed dataset and auxiliary tasks effectively help PLMs to generate humorous responses, laying the groundwork for future humor research.

Graph neural networks (GNNs) are commonly employed in collaborative friend recommendation systems. Nevertheless, recent studies reveal a notable performance gap, particularly for users with limited connections, commonly known as tail users, in contrast to their counterparts with abundant connections (head users). Uniformly treating head and tail users poses two challenges for tail user preference learning: (C1) Label Sparsity, as tail users typically possess limited labels; and (C2) Neighborhood Sparsity, where tail users exhibit sparse observable friendships, leading to distinct preference distributions and performance degradation compared to head users. In response to these challenges, we introduce Tail-STEAK, an innovative framework that combines self-training with enhanced knowledge distillation for tail user representation learning. To address(C1), we present Tail-STEAK-base, a two-stage self-training framework. In the first stage, only head users and their accurate connections are utilized for training, while pseudo links are generated for tail users in the second stage. To tackle (C2), we propose two data augmentation-based self-knowledge distillation pretext tasks. These tasks are seamlessly integrated into different stages of Tail-STEAK-base, culminating in the comprehensive Tail-STEAK framework. Extensive experiments, conducted on state-of-the-art GNN-based friend recommendation models, substantiate the efficacy of Tail-STEAK in significantly improving tail user performance. Our code and data are publicly available at https://github.com/antman9914/Tail-STEAK.

Out-of-distribution (OOD) detection is crucial in many real-world applications. However, intelligent models are often trained solely on in-distribution (ID) data, leading to overconfidence when misclassifying OOD data as ID classes. In this study, we propose a new learning framework which leverage simple Jigsaw-based fake OOD data and rich semantic embeddings (`anchors') from the ChatGPT description of ID knowledge to help guide the training of the image encoder. The learning framework can be flexibly combined with existing post-hoc approaches to OOD detection, and extensive empirical evaluations on multiple OOD detection benchmarks demonstrate that rich textual representation of ID knowledge and fake OOD knowledge can well help train a visual encoder for OOD detection. With the learning framework, new state-of-the-art performance was achieved on all the benchmarks. The code is available at https://github.com/Cverchen/TagFog.

Contrastive Language-Image Pre-training (CLIP) has demonstrated impressive capabilities in open-vocabulary classification. The class token in the image encoder is trained to capture the global features to distinguish different text descriptions supervised by contrastive loss, making it highly effective for single-label classification. However, it shows poor performance on multi-label datasets because the global feature tends to be dominated by the most prominent class and the contrastive nature of softmax operation aggravates it. In this study, we observe that the multi-label classification results heavily rely on discriminative local features but are overlooked by CLIP. As a result, we dissect the preservation of patch-wise spatial information in CLIP and proposed a local-to-global framework to obtain image tags. It comprises three steps: (1) patch-level classification to obtain coarse scores; (2) dual-masking attention refinement (DMAR) module to refine the coarse scores; (3) class-wise reidentification (CWR) module to remedy predictions from a global perspective. This framework is solely based on frozen CLIP and significantly enhances its multi-label classification performance on various benchmarks without dataset-specific training. Besides, to comprehensively assess the quality and practicality of generated tags, we extend their application to the downstream task, i.e., weakly supervised semantic segmentation (WSSS) with generated tags as image-level pseudo labels. Experiments demonstrate that this classify-then-segment paradigm dramatically outperforms other annotation-free segmentation methods and validates the effectiveness of generated tags. Our code is available at https://github.com/linyq2117/TagCLIP.

Visual language tasks such as Visual Question Answering (VQA) or Visual Entailment (VE) require AI models to comprehend and reason with both visual and textual content. Driven by the power of Large Language Models (LLMs), two prominent methods have emerged: (1) the hybrid integration between LLMs and Vision-Language Models (VLMs), where visual inputs are firstly converted into language descriptions by VLMs, serving as inputs for LLMs to generate final answer(s); (2) visual feature alignment in language space, where visual inputs are encoded as embeddings and projected to LLMs' language space via further supervised fine-tuning. The first approach provides light training costs and interpretability but is hard to be optimized in an end-to-end fashion. The second approach presents decent performance, but feature alignment usually requires large amounts of training data and lacks interpretability. To tackle this dilemma, we propose a novel approach, Inner Monologue Multi-Modal Optimization (IMMO), to solve complex vision language problems by simulating Inner Monologue, a cognitive process in which an individual engages in silent verbal communication with themselves. More specifically, we enable LLMs and VLMs to interact through natural language conversation (i.e., Inner Monologue) and propose to use a two-stage training process to learn how to do Inner Monologue (self-asking questions and answering questions). IMMO is evaluated on two popular tasks and achieves competitive performance with less training data when compared with state-of-the-art models while concurrently keeping the interpretability. The results suggest that by emulating the cognitive phenomenon of internal dialogue, our approach can enhance reasoning and explanation abilities, contributing to the more effective fusion of vision and language models. More importantly, instead of using predefined human-crafted monologues, IMMO learns this process within the deep learning models, broadening its potential applications across various AI challenges beyond vision and language tasks.

Decision trees are widely used for addressing learning tasks involving tabular data. Yet, they are susceptible to adversarial attacks. In this paper, we present Tree Test Time Simulation (TTTS), a novel inference-time methodology that incorporates Monte Carlo simulations into decision trees to enhance their robustness. TTTS introduces a probabilistic modification to the decision path, without altering the underlying tree structure. Our comprehensive empirical analysis of 50 datasets yields promising results. Without the presence of any attacks, TTTS has successfully improved model performance from an AUC of 0.714 to 0.773. Under the challenging conditions of white-box attacks, TTTS demonstrated its robustness by boosting performance from an AUC of 0.337 to 0.680. Even when subjected to black-box attacks, TTTS maintains high accuracy and enhances the model's performance from an AUC of 0.628 to 0.719. Compared to defenses such as Feature Squeezing, TTTS proves to be much more effective. We also found that TTTS exhibits similar robustness in decision forest settings across different attacks.

When dealing with tabular data, models based on decision trees are a popular choice due to their high accuracy on these data types, their ease of application, and explainability properties. However, when it comes to graph-structured data, it is not clear how to apply them effectively, in a way that in- corporates the topological information with the tabular data available on the vertices of the graph. To address this challenge, we introduce TREE-G. TREE-G modifies standard decision trees, by introducing a novel split function that is specialized for graph data. Not only does this split function incorporate the node features and the topological information, but it also uses a novel pointer mechanism that allows split nodes to use information computed in previous splits. Therefore, the split function adapts to the predictive task and the graph at hand. We analyze the theoretical properties of TREE-G and demonstrate its benefits empirically on multiple graph and vertex prediction benchmarks. In these experiments, TREE-G consistently outperforms other tree-based models and often outperforms other graph-learning algorithms such as Graph Neural Networks (GNNs) and Graph Kernels, sometimes by large margins. Moreover, TREE-Gs models and their predic tions can be explained and visualized.

Video moment retrieval (MR) and highlight detection (HD) based on natural language queries are two highly related tasks, which aim to obtain relevant moments within videos and highlight scores of each video clip. Recently, several methods have been devoted to building DETR-based networks to solve both MR and HD jointly. These methods simply add two separate task heads after multi-modal feature extraction and feature interaction, achieving good performance. Nevertheless, these approaches underutilize the reciprocal relationship between two tasks. In this paper, we propose a task-reciprocal transformer based on DETR (TR-DETR) that focuses on exploring the inherent reciprocity between MR and HD. Specifically, a local-global multi-modal alignment module is first built to align features from diverse modalities into a shared latent space. Subsequently, a visual feature refinement is designed to eliminate query-irrelevant information from visual features for modal interaction. Finally, a task cooperation module is constructed to refine the retrieval pipeline and the highlight score prediction process by utilizing the reciprocity between MR and HD. Comprehensive experiments on QVHighlights, Charades-STA and TVSum datasets demonstrate that TR-DETR outperforms existing state-of-the-art methods. Codes are available at https://github.com/mingyao1120/TR-DETR.

Multi-spectral object Re-identification (ReID) aims to retrieve specific objects by leveraging complementary information from different image spectra. It delivers great advantages over traditional single-spectral ReID in complex visual environment. However, the significant distribution gap among different image spectra poses great challenges for effective multi-spectral feature representations. In addition, most of current Transformer-based ReID methods only utilize the global feature of class tokens to achieve the holistic retrieval, ignoring the local discriminative ones. To address the above issues, we step further to utilize all the tokens of Transformers and propose a cyclic token permutation framework for multi-spectral object ReID, dubbled TOP-ReID. More specifically, we first deploy a multi-stream deep network based on vision Transformers to preserve distinct information from different image spectra. Then, we propose a Token Permutation Module (TPM) for cyclic multi-spectral feature aggregation. It not only facilitates the spatial feature alignment across different image spectra, but also allows the class token of each spectrum to perceive the local details of other spectra. Meanwhile, we propose a Complementary Reconstruction Module (CRM), which introduces dense token-level reconstruction constraints to reduce the distribution gap across different image spectra. With the above modules, our proposed framework can generate more discriminative multi-spectral features for robust object ReID. Extensive experiments on three ReID benchmarks (i.e., RGBNT201, RGBNT100 and MSVR310) verify the effectiveness of our methods. The code is available at https://github.com/924973292/TOP-ReID.

Learning Granger causality from event sequences is a challenging but essential task across various applications. Most existing methods rely on the assumption that event sequences are independent and identically distributed (i.i.d.). However, this i.i.d. assumption is often violated due to the inherent dependencies among the event sequences. Fortunately, in practice, we find these dependencies can be modeled by a topological network, suggesting a potential solution to the non-i.i.d. problem by introducing the prior topological network into Granger causal discovery. This observation prompts us to tackle two ensuing challenges: 1) how to model the event sequences while incorporating both the prior topological network and the latent Granger causal structure, and 2) how to learn the Granger causal structure. To this end, we devise a unified topological neural Poisson auto-regressive model with two processes. In the generation process, we employ a variant of the neural Poisson process to model the event sequences, considering influences from both the topological network and the Granger causal structure. In the inference process, we formulate an amortized inference algorithm to infer the latent Granger causal structure. We encapsulate these two processes within a unified likelihood function, providing an end-to-end framework for this task. Experiments on simulated and real-world data demonstrate the effectiveness of our approach.

The paper presents Taylor Map Polynomial Neural Network (TMPNN), a novel form of very high-order polynomial regression, in which the same coefficients for a lower-to-moderate-order polynomial regression are iteratively reapplied so as to achieve a higher-order model without the number of coefficients to be fit exploding in the usual curse-of-dimensionality way. This method naturally implements multi-target regression and can capture internal relationships between targets. We also introduce an approach for model interpretation in the form of systems of differential equations. By benchmarking on Feynman regression, UCI, Friedman-1, and real-life industrial datasets, we demonstrate that the proposed method performs comparably to the state-of-the-art regression methods and outperforms them on specific tasks.

Numerous techniques excel in brain tumor segmentation using multi-modal magnetic resonance imaging (MRI) sequences, delivering exceptional results. However, the prevalent absence of modalities in clinical scenarios hampers performance. Current approaches frequently resort to zero maps as substitutes for missing modalities, inadvertently introducing feature bias and redundant computations. To address these issues, we present the Token Merging transFormer (TMFormer) for robust brain tumor segmentation with missing modalities. TMFormer tackles these challenges by extracting and merging accessible modalities into more compact token sequences. The architecture comprises two core components: the Uni-modal Token Merging Block (UMB) and the Multi-modal Token Merging Block (MMB). The UMB enhances individual modality representation by adaptively consolidating spatially redundant tokens within and outside tumor-related regions, thereby refining token sequences for augmented representational capacity. Meanwhile, the MMB mitigates multi-modal feature fusion bias, exclusively leveraging tokens from present modalities and merging them into a unified multi-modal representation to accommodate varying modality combinations. Extensive experimental results on the BraTS 2018 and 2020 datasets demonstrate the superiority and efficacy of TMFormer compared to state-of-the-art methods when dealing with missing modalities.

Continual Semantic Segmentation (CSS) is an emerging trend, where catastrophic forgetting has been a perplexing problem. In this paper, we propose a Text-to-Image Knowledge Preservation (TIKP) framework to address this issue. TIKP applies Text-to-Image techniques to CSS by automatically generating prompts and content adaptation. It extracts associations between the labels of seen data and constructs text-level prompts based on these associations, which are preserved and maintained at each incremental step. During training, these prompts generate correlated images to mitigate the catastrophic forgetting. Particularly, as the generated images may have different distributions from the original data, TIKP transfers the knowledge by a content adaption loss, which determines the role played by the generated images in incremental training based on the similarity. In addition, for the classifier, we use the previous model from a different perspective: misclassifying new classes into old objects instead of the background. We propose a knowledge distillation loss based on wrong labels, enabling us to attribute varying weights to individual objects during the distillation process. Extensive experiments conducted in the same setting show that TIKP outperforms state-of-the-art methods by a large margin on benchmark datasets.