AAAI 2024 Papers

This work presents an effective depth-consistency Self-Prompt Transformer, terms as SelfPromer, for image dehazing. It is motivated by an observation that the estimated depths of an image with haze residuals and its clear counterpart vary. Enforcing the depth consistency of dehazed images with clear ones, therefore, is essential for dehazing. For this purpose, we develop a prompt based on the features of depth differences between the hazy input images and corresponding clear counterparts that can guide dehazing models for better restoration. Specifically, we first apply deep features extracted from the input images to the depth difference features for generating the prompt that contains the haze residual information in the input. Then we propose a prompt embedding module that is designed to perceive the haze residuals, by linearly adding the prompt to the deep features. Further, we develop an effective prompt attention module to pay more attention to haze residuals for better removal. By incorporating the prompt, prompt embedding, and prompt attention into an encoder-decoder network based on VQGAN, we can achieve better perception quality. As the depths of clear images are not available at inference, and the dehazed images with one-time feed-forward execution may still contain a portion of haze residuals, we propose a new continuous self-prompt inference that can iteratively correct the dehazing model towards better haze-free image generation. Extensive experiments show that our SelfPromer performs favorably against the state-of-the-art approaches on both synthetic and real-world datasets in terms of perception metrics including NIQE, PI, and PIQE. The source codes will be made available at https://github.com/supersupercong/SelfPromer.

Self-training based few-shot node classification (FSNC) methods have shown excellent performance in real applications, but they cannot make the full use of the information in the base set and are easily affected by the quality of pseudo-labels. To address these issues, this paper proposes a new self-training FSNC method by involving the representation distillation and the pseudo-label distillation. Specifically, the representation distillation includes two knowledge distillation methods (i.e., the local representation distillation and the global representation distillation) to transfer the information in the base set to the novel set. The pseudo-label distillation is designed to conduct knowledge distillation on the pseudo-labels to improve their quality. Experimental results showed that our method achieves supreme performance, compared with state-of-the-art methods. Our code and a comprehensive theoretical version are available at https://github.com/zongqianwu/KD-FSNC.

Self-Supervised Learning (SSL) methods harness the concept of semantic invariance by utilizing data augmentation strategies to produce similar representations for different deformations of the same input. Essentially, the model captures the shared information among multiple augmented views of samples, while disregarding the non-shared information that may be beneficial for downstream tasks. To address this issue, we introduce a module called CompMod with Meta Comprehensive Regularization (MCR), embedded into existing self-supervised frameworks, to make the learned representations more comprehensive. Specifically, we update our proposed model through a bi-level optimization mechanism, enabling it to capture comprehensive features. Additionally, guided by the constrained extraction of features using maximum entropy coding, the self-supervised learning model learns more comprehensive features on top of learning consistent features. In addition, we provide theoretical support for our proposed method from information theory and causal counterfactual perspective. Experimental results show that our method achieves significant improvement in classification, object detection and semantic segmentation tasks on multiple benchmark datasets.

Deep cross-modal hashing technology provides an effective and efficient cross-modal unified representation learning solution for cross-modal search. However, the existing methods neglect the implicit fine-grained multimodal knowledge relations between these modalities such as when the image contains information that is not directly described in the text. To tackle this problem, we propose a novel self-supervised multi-grained multi-modal knowledge graph contrastive hashing method for cross-modal search (CMGCH). Firstly, in order to capture implicit fine-grained cross-modal semantic associations, a multi-modal knowledge graph is constructed, which represents the implicit multimodal knowledge relations between the image and text as inter-modal and intra-modal semantic associations. Secondly, a cross-modal graph contrastive attention network is proposed to reason on the multi-modal knowledge graph to sufficiently learn the implicit fine-grained inter-modal and intra-modal knowledge relations. Thirdly, a cross-modal multi-granularity contrastive embedding learning mechanism is proposed, which fuses the global coarse-grained and local fine-grained embeddings by multihead attention mechanism for inter-modal and intra-modal contrastive learning, so as to enhance the cross-modal unified representations with stronger discriminativeness and semantic consistency preserving power. With the joint training of intra-modal and inter-modal contrast, the invariant and modal-specific information of different modalities can be maintained in the final unified cross-modal unified hash space. Extensive experiments on several cross-modal benchmark datasets demonstrate that the proposed CMGCH outperforms the state-of the-art methods.

Robust autonomous driving requires agents to accurately identify unexpected areas (anomalies) in urban scenes. To this end, some critical issues remain open: how to design advisable metric to measure anomalies, and how to properly generate training samples of anomaly data? Classical effort in anomaly detection usually resorts to pixel-wise uncertainty or sample synthesis, which ignores the contextual information and sometimes requires auxiliary data with fine-grained annotations. On the contrary, in this paper, we exploit the strong context-dependent nature of segmentation task and design an energy-guided self-supervised frameworks for anomaly segmentation, which optimizes an anomaly head by maximizing likelihood of self-generated anomaly pixels. For this purpose, we design two estimators to model anomaly likelihood, one is a task-agnostic binary estimator and the other depicts the likelihood as residual of task-oriented joint energy. Based on proposed estimators, we devise an adaptive self-supervised training framework, which exploits the contextual reliance and estimated likelihood to refine mask annotations in anomaly areas. We conduct extensive experiments on challenging Fishyscapes and Road Anomaly benchmarks, demonstrating that without any auxiliary data or synthetic models, our method can still achieves comparable performance to supervised competitors. Code is available at https://github.com/yuanpengtu/SLEEG.

With the widespread adoption of IoT, wearable devices, and sensors, time series data from human subjects are significantly increasing in the healthcare domain. Due to the laborious nature of manual annotation in time series data and the requirement for human experts, self-supervised learning methods are attempted to alleviate the limited label situations. While existing self-supervised methods have been successful to achieve comparable performance to the fully supervised methods, there are still some limitations that need to be addressed, considering the nature of time series data from human subjects: In real-world clinical settings, data labels (e.g., sleep stages) are usually annotated by subject-level, and there is a substantial variation in patterns between subjects. Thus, a model should be designed to deal with not only the label scarcity but also subject-wise nature of data to ensure high performance in real-world scenarios. To mitigate these issues, we propose a novel self-supervised learning framework for human subject time series data: Subject-Aware Time Series Clustering (SA-TSC). In the unsupervised representation learning phase, SA-TSC adopts a subject-wise learning strategy rather than instance-wise learning which randomly samples data instances from different subjects within the batch during training. Specifically, we generate subject-graphs with our graph construction method based on Gumbel-Softmax and perform graph spectral clustering on each subject-graph. In addition, we utilize graph neural networks to capture dependencies between channels and design our own graph learning module motivated from self-supervised loss. Experimental results show the outstanding performance of our SA-TSC with the limited & subject-wise label setting, leading to its high applicability to the healthcare industry. The code is available at: https://github.com/DILAB-HYU/SA-TSC

Speech signals are inherently complex as they encompass both global acoustic characteristics and local semantic information. However, in the task of target speech extraction, certain elements of global and local semantic information in the reference speech, which are irrelevant to speaker identity, can lead to speaker confusion within the speech extraction network. To overcome this challenge, we propose a self-supervised disentangled representation learning method. Our approach tackles this issue through a two-phase process, utilizing a reference speech encoding network and a global information disentanglement network to gradually disentangle the speaker identity information from other irrelevant factors. We exclusively employ the disentangled speaker identity information to guide the speech extraction network. Moreover, we introduce the adaptive modulation Transformer to ensure that the acoustic representation of the mixed signal remains undisturbed by the speaker embeddings. This component incorporates speaker embeddings as conditional information, facilitating natural and efficient guidance for the speech extraction network. Experimental results substantiate the effectiveness of our meticulously crafted approach, showcasing a substantial reduction in the likelihood of speaker confusion.

Learning the dense bird's eye view (BEV) motion flow in a self-supervised manner is an emerging research for robotics and autonomous driving. Current self-supervised methods mainly rely on point correspondences between point clouds, which may introduce the problems of fake flow and inconsistency, hindering the model’s ability to learn accurate and realistic motion. In this paper, we introduce a novel cross-modality self-supervised training framework that effectively addresses these issues by leveraging multi-modality data to obtain supervision signals. We design three innovative supervision signals to preserve the inherent properties of scene motion, including the masked Chamfer distance loss, the piecewise rigidity loss, and the temporal consistency loss. Through extensive experiments, we demonstrate that our proposed self-supervised framework outperforms all previous self-supervision methods for the motion prediction task.

Despite achieving impressive improvement in accuracy, most existing monocular 3D human mesh reconstruction methods require large-scale 2D/3D ground-truths for supervision, which limits their applications on unlabeled in-the-wild data that is ubiquitous. To alleviate the reliance on 2D/3D ground-truths, we present a self-supervised 3D human pose and shape reconstruction framework that relies only on self-consistency between intermediate representations of images and projected 2D predictions. Specifically, we extract 2D joints and depth maps from monocular images as proxy inputs, which provides complementary clues to infer accurate 3D human meshes. Furthermore, to reduce the impacts from noisy and ambiguous inputs while better concentrate on the high-quality information, we design an uncertainty-aware module to automatically learn the reliability of the inputs at body-joint level based on the consistency between 2D joints and depth map. Experiments on benchmark datasets show that our approach outperforms other state-of-the-art methods at similar supervision levels.

Few-shot learning poses a formidable challenge as it necessitates effective recognition of novel classes based on a limited set of examples. Recent studies have sought to address the challenge of rare samples by tuning visual features through the utilization of external text prompts. However, the performance of these methods is constrained due to the inherent modality gap between the prompt text and image features. Instead of naively utilizing the external semantic information generated from text to guide the training of the image encoder, we propose a novel self-prompt mechanism (SPM) to adaptively adjust the neural network according to unseen data. Specifically, SPM involves a systematic selection of intrinsic semantic features generated by the image encoder across spatial and channel dimensions, thereby engendering self-prompt information. Subsequently, upon backpropagation of this self-prompt information to the deeper layers of the neural network, it effectively steers the network toward the learning and adaptation of new samples. Meanwhile, we propose a novel parameter-efficient tuning method that exclusively fine-tunes the parameters relevant to self-prompt (prompts are no more than 2% of the total parameters), and the incorporation of additional learnable parameters as self-prompt ensures the retention of prior knowledge through frozen encoder weights. Therefore, our method is highly suited for few-shot recognition tasks that require both information retention and adaptive adjustment of network parameters with limited labeling data constraints. Extensive experiments demonstrate the effectiveness of the proposed SPM in both 5-way 1-shot and 5-way 5-shot settings for standard single-domain and cross-domain few-shot recognition datasets, respectively. Our code is available at https://github.com/codeshop715/SPM.

Hierarchical Multi-Label Classification (HMLC) is a well-established problem that aims at assigning data instances to multiple classes stored in a hierarchical structure. Despite its importance, existing approaches often face two key limitations: (i) They employ dense networks to solely explore the class hierarchy as hard criterion for maintaining taxonomic consistency among predicted classes, yet without leveraging rich semantic relationships between instances and classes; (ii) They struggle to generalize in settings with deep class levels, since the mini-batches uniformly sampled from different levels ignore the varying complexities of data and result in a non-smooth model adaptation to sparse data. To mitigate these issues, we present a Self-Paced Unified Representation (SPUR) learning framework, which focuses on the interplay between instance and classes to flexibly organize the training process of HMLC algorithms. Our framework consists of two lightweight encoders designed to capture the semantics of input features and the topological information of the class hierarchy. These encoders generate unified embeddings of instances and class hierarchy, which enable SPUR to exploit semantic dependencies between them and produce predictions in line with taxonomic constraints. Furthermore, we introduce a dynamic hardness measurement strategy that considers both class hierarchy and instance features to estimate the learning difficulty of each instance. This strategy is achieved by incorporating the propagation loss obtained at each hierarchical level, allowing for a more comprehensive assessment of learning complexity. Extensive experiments on several empirical benchmarks demonstrate the effectiveness and efficiency of SPUR compared to state-of-the-art methods, especially in scenarios with missing features.

Self-interpretable graph learning methods provide insights to unveil the black-box nature of GNNs by providing predictions with built-in explanations. However, current works suffer from performance degradation compared to GNNs trained without built-in explanations. We argue the main reason is that they fail to generate explanations satisfying both sufficiency and necessity, and the biased explanations further hurt GNNs' performance. In this work, we propose a novel framework for generating SUfficient aNd NecessarY explanations (SUNNY-GNN for short) that benefit GNNs' predictions. The key idea is to conduct augmentations by structurally perturbing given explanations and employ a contrastive loss to guide the learning of explanations toward sufficiency and necessity directions. SUNNY-GNN introduces two coefficients to generate hard and reliable contrastive samples. We further extend SUNNY-GNN to heterogeneous graphs. Empirical results on various GNNs and real-world graphs show that SUNNY-GNN yields accurate predictions and faithful explanations, outperforming the state-of-the-art methods by improving 3.5% prediction accuracy and 13.1% explainability fidelity on average. Our code and data are available at https://github.com/SJTU-Quant/SUNNY-GNN.

Text recognition methods are gaining rapid development. Some advanced techniques, e.g., powerful modules, language models, and un- and semi-supervised learning schemes, consecutively push the performance on public benchmarks forward. However, the problem of how to better optimize a text recognition model from the perspective of loss functions is largely overlooked. CTC-based methods, widely used in practice due to their good balance between performance and inference speed, still grapple with accuracy degradation. This is because CTC loss emphasizes the optimization of the entire sequence target while neglecting to learn individual characters. We propose a self-distillation scheme for CTC-based model to address this issue. It incorporates a framewise regularization term in CTC loss to emphasize individual supervision, and leverages the maximizing-a-posteriori of latent alignment to solve the inconsistency problem that arises in distillation between CTC-based models. We refer to the regularized CTC loss as Distillation Connectionist Temporal Classification (DCTC) loss. DCTC loss is module-free, requiring no extra parameters, longer inference lag, or additional training data or phases. Extensive experiments on public benchmarks demonstrate that DCTC can boost text recognition model accuracy by up to 2.6%, without any of these drawbacks.

Pedestrian Attribute Recognition (PAR) involves identifying the attributes of individuals in person images. Existing PAR methods typically rely on CNNs as the backbone network to extract pedestrian features. However, CNNs process only one adjacent region at a time, leading to the loss of long-range inter-relations between different attribute-specific regions. To address this limitation, we leverage the Vision Transformer (ViT) instead of CNNs as the backbone for PAR, aiming to model long-range relations and extract more robust features. However, PAR suffers from an inherent attribute imbalance issue, causing ViT to naturally focus more on attributes that appear frequently in the training set and ignore some pedestrian attributes that appear less. The native features extracted by ViT are not able to tolerate the imbalance attribute distribution issue. To tackle this issue, we propose two novel components: the Selective Feature Activation Method (SFAM) and the Orthogonal Feature Activation Loss. SFAM smartly suppresses the more informative attribute-specific features, compelling the PAR model to capture discriminative features from regions that are easily overlooked. The proposed loss enforces an orthogonal constraint on the original feature extracted by ViT and the suppressed features from SFAM, promoting the complementarity of features in space. We conduct experiments on several benchmark PAR datasets, including PETA, PA100K, RAPv1, and RAPv2, demonstrating the effectiveness of our method. Specifically, our method outperforms existing state-of-the-art approaches by GRL, IAA-Caps, ALM, and SSC in terms of mA on the four datasets, respectively.

Lane detection (LD) plays a crucial role in enhancing the L2+ capabilities of autonomous driving, capturing widespread attention. The Post-Processing Quantization (PTQ) could facilitate the practical application of LD models, enabling fast speeds and limited memories without labeled data. However, prior PTQ methods do not consider the complex LD outputs that contain physical semantics, such as offsets, locations, etc., and thus cannot be directly applied to LD models. In this paper, we pioneeringly investigate semantic sensitivity to post-processing for lane detection with a novel Lane Distortion Score. Moreover, we identify two main factors impacting the LD performance after quantization, namely intra-head sensitivity and inter-head sensitivity, where a small quantization error in specific semantics can cause significant lane distortion. Thus, we propose a Selective Focus framework deployed with Semantic Guided Focus and Sensitivity Aware Selection modules, to incorporate post-processing information into PTQ reconstruction. Based on the observed intra-head sensitivity, Semantic Guided Focus is introduced to prioritize foreground-related semantics using a practical proxy. For inter-head sensitivity, we present Sensitivity Aware Selection, efficiently recognizing influential prediction heads and refining the optimization objectives at runtime. Extensive experiments have been done on a wide variety of models including keypoint-, anchor-, curve-, and segmentation-based ones. Our method produces quantized models in minutes on a single GPU and can achieve 6.4\% F1 Score improvement on the CULane dataset. Code and supplementary statement can be found at https://github.com/PannenetsF/SelectiveFocus.

In light of the advances in big data, high-dimensional datasets are often encountered. Incorporating them into data-driven models can enhance performance; however, this comes at the cost of high computation and the risk of overfitting, particularly due to abundant redundant features. Identifying an informative subset of the features helps in reducing the dimensionality and enhancing model interpretability. In this paper, we propose a novel framework for unsupervised feature selection, called Selective Deep Auto-Encoder (SDAE). It aims to reduce the number of features used in unlabeled datasets without compromising the quality of information obtained. It achieves this by selecting sufficient features - from the original feature set - capable of representing the entire feature space and reconstructing them. Architecturally, it leverages the use of highly nonlinear latent representations in deep Autoencoders and intrinsically learns, in an unsupervised fashion, the relevant and globally representative subset of features through a customized Selective Layer. Extensive experimental results on three high-dimensional public datasets have shown promising feature selection performance by SDAE in comparison to other existing state-of-the-art unsupervised feature selection methods.

Injecting textual information into knowledge graph (KG) entity representations has been a worthwhile expedition in terms of improving performance in KG oriented tasks within the NLP community. External knowledge often adopted to enhance KG embeddings ranges from semantically rich lexical dependency parsed features to a set of relevant key words to entire text descriptions supplied from an external corpus such as wikipedia and many more. Despite the gains this innovation (Text-enhanced KG embeddings) has made, the proposal in this work suggests that it can be improved even further. Instead of using a single text description (which would not sufficiently represent an entity because of the inherent lexical ambiguity of text), we propose a multi-task framework that jointly selects a set of text descriptions relevant to KG entities as well as align or augment KG embeddings with text descriptions. Different from prior work that plugs formal entity descriptions declared in knowledge bases, this framework leverages a retriever model to selectively identify richer or highly relevant text descriptions to use in augmenting entities. Furthermore, the framework treats the number of descriptions to use in augmentation process as a parameter, which allows the flexibility of enumerating across several numbers before identifying an appropriate number. Experiment results for Link Prediction demonstrate a 5.5% and 3.5% percentage increase in the Mean Reciprocal Rank (MRR) and Hits@10 scores respectively, in comparison to text-enhanced knowledge graph augmentation methods using traditional CNNs.

Augmented Reality (AR) devices, emerging as prominent mobile interaction platforms, face challenges in user safety, particularly concerning oncoming vehicles. While some solutions leverage onboard camera arrays, these cameras often have limited field-of-view (FoV) with front or downward perspectives. Addressing this, we propose a new out-of-view semantic segmentation task and Segment Beyond View (SBV), a novel audio-visual semantic segmentation method. SBV supplements the visual modality, which miss the information beyond FoV, with the auditory information using a teacher-student distillation model (Omni2Ego). The model consists of a vision teacher utilising panoramic information, an auditory teacher with 8-channel audio, and an audio-visual student that takes views with limited FoV and binaural audio as input and produce semantic segmentation for objects outside FoV. SBV outperforms existing models in comparative evaluations and shows a consistent performance across varying FoV ranges and in monaural audio settings.

Enhancing low-light videos in a supervised style presents a set of challenges, including limited data diversity, misalignment, and the domain gap introduced through the dataset construction pipeline. Our paper tackles these challenges by constructing a self-learned enhancement approach that gets rid of the reliance on any external training data. The challenge of self-supervised learning lies in fitting high-quality signal representations solely from input signals. Our work designs a bottleneck neural representation mechanism that extracts those signals. More in detail, we encode the frame-wise representation with a compact deep embedding and utilize a neural network to parameterize the video-level manifold consistently. Then, an entropy constraint is applied to the enhanced results based on the adjacent spatial-temporal context to filter out the degraded visual signals, e.g. noise and frame inconsistency. Last, a novel Chromatic Retinex decomposition is proposed to effectively align the reflectance distribution temporally. It benefits the entropy control on different components of each frame and facilitates noise-to-noise training, successfully suppressing the temporal flicker. Extensive experiments demonstrate the robustness and superior effectiveness of our proposed method. Our project is publicly available at: https://huangerbai.github.io/SLBNR/.