This work investigates how hierarchically structured data can help neural networks learn conceptual representations of cathedrals. The underlying WikiScenes dataset provides a spatially organized hierarchical structure of cathedral components. We propose a novel hierarchical contrastive training approach that leverages a triplet margin loss to represent the data's spatial hierarchy in the encoder's latent space. As such, the proposed approach investigates if the dataset structure provides valuable information for self-supervised learning. We apply t-SNE to visualize the resultant latent space and evaluate the proposed approach by comparing it with other dataset-specific contrastive learning methods using a common downstream classification task. The proposed method outperforms the comparable weakly-supervised and baseline methods. Our findings suggest that dataset structure is a valuable modality for weakly-supervised learning.
In recent years, there has been growing interest in the video-based action quality assessment (AQA). Most existing methods typically solve AQA problem by considering the entire video yet overlooking the inherent stage-level characteristics of actions. To address this issue, we design a novel Multi-stage Contrastive Regression (MCoRe) framework for the AQA task. This approach allows us to efficiently extract spatial-temporal information, while simultaneously reducing computational costs by segmenting the input video into multiple stages or procedures. Inspired by the graph contrastive learning, we propose a new stage-wise contrastive learning loss function to enhance performance. As a result, MCoRe demonstrates the state-of-the-art result so far on the widely-adopted fine-grained AQA dataset.
Self-supervised frameworks for representation learning have recently stirred up interest among the remote sensing community, given their potential to mitigate the high labeling costs associated with curating large satellite image datasets. In the realm of multimodal data fusion, while the often used contrastive learning methods can help bridging the domain gap between different sensor types, they rely on data augmentations techniques that require expertise and careful design, especially for multispectral remote sensing data. A possible but rather scarcely studied way to circumvent these limitations is to use a masked image modelling based pretraining strategy. In this paper, we introduce Fus-MAE, a self-supervised learning framework based on masked autoencoders that uses cross-attention to perform early and feature-level data fusion between synthetic aperture radar and multispectral optical data - two modalities with a significant domain gap. Our empirical findings demonstrate that Fus-MAE can effectively compete with contrastive learning strategies tailored for SAR-optical data fusion and outperforms other masked-autoencoders frameworks trained on a larger corpus.
We present Unsupervised hard Negative Augmentation (UNA), a method that generates synthetic negative instances based on the term frequency-inverse document frequency (TF-IDF) retrieval model. UNA uses TF-IDF scores to ascertain the perceived importance of terms in a sentence and then produces negative samples by replacing terms with respect to that. Our experiments demonstrate that models trained with UNA improve the overall performance in semantic textual similarity tasks. Additional performance gains are obtained when combining UNA with the paraphrasing augmentation. Further results show that our method is compatible with different backbone models. Ablation studies also support the choice of having a TF-IDF-driven control on negative augmentation.
This paper focuses on learning representation on the whole graph level in an unsupervised manner. Learning graph-level representation plays an important role in a variety of real-world issues such as molecule property prediction, protein structure feature extraction, and social network analysis. The mainstream method is utilizing contrastive learning to facilitate graph feature extraction, known as Graph Contrastive Learning (GCL). GCL, although effective, suffers from some complications in contrastive learning, such as the effect of false negative pairs. Moreover, augmentation strategies in GCL are weakly adaptive to diverse graph datasets. Motivated by these problems, we propose a novel framework called Structure Knowledge Refinement (SKR) which uses data structure to determine the probability of whether a pair is positive or negative. Meanwhile, we propose an augmentation strategy that naturally preserves the semantic meaning of the original data and is compatible with our SKR framework. Furthermore, we illustrate the effectiveness of our SKR framework through intuition and experiments. The experimental results on the tasks of graph-level classification demonstrate that our SKR framework is superior to most state-of-the-art baselines.
While word error rates of automatic speech recognition (ASR) systems have consistently fallen, natural language understanding (NLU) applications built on top of ASR systems still attribute significant numbers of failures to low-quality speech recognition results. Existing assistant systems collect large numbers of these unsuccessful interactions, but these systems usually fail to learn from these interactions, even in an offline fashion. In this work, we introduce CLC: Contrastive Learning for Conversations, a family of methods for contrastive fine-tuning of models in a self-supervised fashion, making use of easily detectable artifacts in unsuccessful conversations with assistants. We demonstrate that our CLC family of approaches can improve the performance of ASR models on OD3, a new public large-scale semi-synthetic meta-dataset of audio task-oriented dialogues, by up to 19.2%. These gains transfer to real-world systems as well, where we show that CLC can help to improve performance by up to 6.7% over baselines. We make OD3 publicly available at https://github.com/amazon-science/amazon-od3 .
Wearable-based Human Activity Recognition (HAR) is a key task in human-centric machine learning due to its fundamental understanding of human behaviours. Due to the dynamic nature of human behaviours, continual learning promises HAR systems that are tailored to users' needs. However, because of the difficulty in collecting labelled data with wearable sensors, existing approaches that focus on supervised continual learning have limited applicability, while unsupervised continual learning methods only handle representation learning while delaying classifier training to a later stage. This work explores the adoption and adaptation of CaSSLe, a continual self-supervised learning model, and Kaizen, a semi-supervised continual learning model that balances representation learning and down-stream classification, for the task of wearable-based HAR. These schemes re-purpose contrastive learning for knowledge retention and, Kaizen combines that with self-training in a unified scheme that can leverage unlabelled and labelled data for continual learning. In addition to comparing state-of-the-art self-supervised continual learning schemes, we further investigated the importance of different loss terms and explored the trade-off between knowledge retention and learning from new tasks. In particular, our extensive evaluation demonstrated that the use of a weighting factor that reflects the ratio between learned and new classes achieves the best overall trade-off in continual learning.
Defining pathologies automatically from medical images aids the understanding of the emergence and progression of diseases, and such an ability is crucial in clinical diagnostics. However, existing deep learning models heavily rely on expert annotations and lack generalization capabilities in open clinical environments. In this study, we present a generalizable vision-language model for Annotation-Free pathology Localization (AFLoc). The core strength of AFLoc lies in its extensive multi-level semantic structure-based contrastive learning, which comprehensively aligns multi-granularity medical concepts from reports with abundant image features, to adapt to the diverse expressions of pathologies and unseen pathologies without the reliance on image annotations from experts. We conducted primary experiments on a dataset of 220K pairs of image-report chest X-ray images, and performed extensive validation across six external datasets encompassing 20 types of chest pathologies. The results demonstrate that AFLoc outperforms state-of-the-art methods in both annotation-free localization and classification tasks. Additionally, we assessed the generalizability of AFLoc on other modalities, including histopathology and retinal fundus images. Extensive experiments show that AFLoc exhibits robust generalization capabilities, even surpassing human benchmarks in localizing five different types of pathological images. These results highlight the potential of AFLoc in reducing annotation requirements and its applicability in complex clinical environments.
In recent years, foundation models (FMs) have solidified their role as cornerstone advancements in the deep learning domain. By extracting intricate patterns from vast datasets, these models consistently achieve state-of-the-art results across a spectrum of downstream tasks, all without necessitating extensive computational resources. Notably, MedCLIP, a vision-language contrastive learning-based medical FM, has been designed using unpaired image-text training. While the medical domain has often adopted unpaired training to amplify data, the exploration of potential security concerns linked to this approach hasn't kept pace with its practical usage. Notably, the augmentation capabilities inherent in unpaired training also indicate that minor label discrepancies can result in significant model deviations. In this study, we frame this label discrepancy as a backdoor attack problem. We further analyze its impact on medical FMs throughout the FM supply chain. Our evaluation primarily revolves around MedCLIP, emblematic of medical FM employing the unpaired strategy. We begin with an exploration of vulnerabilities in MedCLIP stemming from unpaired image-text matching, termed BadMatch. BadMatch is achieved using a modest set of wrongly labeled data. Subsequently, we disrupt MedCLIP's contrastive learning through BadDist-assisted BadMatch by introducing a Bad-Distance between the embeddings of clean and poisoned data. Additionally, combined with BadMatch and BadDist, the attacking pipeline consistently fends off backdoor assaults across diverse model designs, datasets, and triggers. Also, our findings reveal that current defense strategies are insufficient in detecting these latent threats in medical FMs' supply chains.
Anomaly detection on attributed networks aims to find the nodes whose behaviors are significantly different from other majority nodes. Generally, network data contains information about relationships between entities, and the anomaly is usually embodied in these relationships. Therefore, how to comprehensively model complex interaction patterns in networks is still a major focus. It can be observed that anomalies in networks violate the homophily assumption. However, most existing studies only considered this phenomenon obliquely rather than explicitly. Besides, the node representation of normal entities can be perturbed easily by the noise relationships introduced by anomalous nodes. To address the above issues, we present a novel contrastive learning framework for anomaly detection on attributed networks, \textbf{SCALA}, aiming to improve the embedding quality of the network and provide a new measurement of qualifying the anomaly score for each node by introducing sparsification into the conventional method. Extensive experiments are conducted on five benchmark real-world datasets and the results show that SCALA consistently outperforms all baseline methods significantly.
Existing contrastive language-image pre-training aims to learn a joint representation by matching abundant image-text pairs. However, the number of image-text pairs in medical datasets is usually orders of magnitude smaller than that in natural datasets. Besides, medical image-text pairs often involve numerous complex fine-grained correspondences. This paper aims to enhance the data efficiency by introducing multiple-to-multiple local relationship modeling to capture denser supervisions. More specifically, we propose a Medical Language-Image Pre-training (MLIP) framework, which exploits the limited image-text medical data more efficiently through patch-sentence matching. Furthermore, we introduce a masked contrastive learning strategy with semantic integrity estimation to reduce redundancy in images while preserving the underlying semantics. Our evaluation results show that MLIP outperforms previous work in zero/few-shot classification and few-shot segmentation tasks by a large margin.
Multimodal deep learning utilizing imaging and diagnostic reports has made impressive progress in the field of medical imaging diagnostics, demonstrating a particularly strong capability for auxiliary diagnosis in cases where sufficient annotation information is lacking. Nonetheless, localizing diseases accurately without detailed positional annotations remains a challenge. Although existing methods have attempted to utilize local information to achieve fine-grained semantic alignment, their capability in extracting the fine-grained semantics of the comprehensive context within reports is limited. To address this problem, a new method is introduced that takes full sentences from textual reports as the basic units for local semantic alignment. This approach combines chest X-ray images with their corresponding textual reports, performing contrastive learning at both global and local levels. The leading results obtained by this method on multiple datasets confirm its efficacy in the task of lesion localization.
In terms of human-computer interaction, it is becoming more and more important to correctly understand the user's emotional state in a conversation, so the task of multimodal emotion recognition (MER) started to receive more attention. However, existing emotion classification methods usually perform classification only once. Sentences are likely to be misclassified in a single round of classification. Previous work usually ignores the similarities and differences between different morphological features in the fusion process. To address the above issues, we propose a two-stage emotion recognition model based on graph contrastive learning (TS-GCL). First, we encode the original dataset with different preprocessing modalities. Second, a graph contrastive learning (GCL) strategy is introduced for these three modal data with other structures to learn similarities and differences within and between modalities. Finally, we use MLP twice to achieve the final emotion classification. This staged classification method can help the model to better focus on different levels of emotional information, thereby improving the performance of the model. Extensive experiments show that TS-GCL has superior performance on IEMOCAP and MELD datasets compared with previous methods.
Multi-label image classification presents a challenging task in many domains, including computer vision and medical imaging. Recent advancements have introduced graph-based and transformer-based methods to improve performance and capture label dependencies. However, these methods often include complex modules that entail heavy computation and lack interpretability. In this paper, we propose Probabilistic Multi-label Contrastive Learning (ProbMCL), a novel framework to address these challenges in multi-label image classification tasks. Our simple yet effective approach employs supervised contrastive learning, in which samples that share enough labels with an anchor image based on a decision threshold are introduced as a positive set. This structure captures label dependencies by pulling positive pair embeddings together and pushing away negative samples that fall below the threshold. We enhance representation learning by incorporating a mixture density network into contrastive learning and generating Gaussian mixture distributions to explore the epistemic uncertainty of the feature encoder. We validate the effectiveness of our framework through experimentation with datasets from the computer vision and medical imaging domains. Our method outperforms the existing state-of-the-art methods while achieving a low computational footprint on both datasets. Visualization analyses also demonstrate that ProbMCL-learned classifiers maintain a meaningful semantic topology.
In open-domain Question Answering (QA), dense retrieval is crucial for finding relevant passages for answer generation. Typically, contrastive learning is used to train a retrieval model that maps passages and queries to the same semantic space. The objective is to make similar ones closer and dissimilar ones further apart. However, training such a system is challenging due to the false negative issue, where relevant passages may be missed during data annotation. Hard negative sampling, which is commonly used to improve contrastive learning, can introduce more noise in training. This is because hard negatives are those closer to a given query, and thus more likely to be false negatives. To address this issue, we propose a novel contrastive confidence regularizer for Noise Contrastive Estimation (NCE) loss, a commonly used loss for dense retrieval. Our analysis shows that the regularizer helps dense retrieval models be more robust against false negatives with a theoretical guarantee. Additionally, we propose a model-agnostic method to filter out noisy negative passages in the dataset, improving any downstream dense retrieval models. Through experiments on three datasets, we demonstrate that our method achieves better retrieval performance in comparison to existing state-of-the-art dense retrieval systems.
The sequential interaction network usually find itself in a variety of applications, e.g., recommender system. Herein, inferring future interaction is of fundamental importance, and previous efforts are mainly focused on the dynamics in the classic zero-curvature Euclidean space. Despite the promising results achieved by previous methods, a range of significant issues still largely remains open: On the bipartite nature, is it appropriate to place user and item nodes in one identical space regardless of their inherent difference? On the network dynamics, instead of a fixed curvature space, will the representation spaces evolve when new interactions arrive continuously? On the learning paradigm, can we get rid of the label information costly to acquire? To address the aforementioned issues, we propose a novel Contrastive model for Sequential Interaction Network learning on Co-Evolving RiEmannian spaces, CSINCERE. To the best of our knowledge, we are the first to introduce a couple of co-evolving representation spaces, rather than a single or static space, and propose a co-contrastive learning for the sequential interaction network. In CSINCERE, we formulate a Cross-Space Aggregation for message-passing across representation spaces of different Riemannian geometries, and design a Neural Curvature Estimator based on Ricci curvatures for modeling the space evolvement over time. Thereafter, we present a Reweighed Co-Contrast between the temporal views of the sequential network, so that the couple of Riemannian spaces interact with each other for the interaction prediction without labels. Empirical results on 5 public datasets show the superiority of CSINCERE over the state-of-the-art methods.
Graphs are typical non-Euclidean data of complex structures. In recent years, Riemannian graph representation learning has emerged as an exciting alternative to Euclidean ones. However, Riemannian methods are still in an early stage: most of them present a single curvature (radius) regardless of structural complexity, suffer from numerical instability due to the exponential/logarithmic map, and lack the ability to capture motif regularity. In light of the issues above, we propose the problem of \emph{Motif-aware Riemannian Graph Representation Learning}, seeking a numerically stable encoder to capture motif regularity in a diverse-curvature manifold without labels. To this end, we present a novel Motif-aware Riemannian model with Generative-Contrastive learning (MotifRGC), which conducts a minmax game in Riemannian manifold in a self-supervised manner. First, we propose a new type of Riemannian GCN (D-GCN), in which we construct a diverse-curvature manifold by a product layer with the diversified factor, and replace the exponential/logarithmic map by a stable kernel layer. Second, we introduce a motif-aware Riemannian generative-contrastive learning to capture motif regularity in the constructed manifold and learn motif-aware node representation without external labels. Empirical results show the superiority of MofitRGC.
Recent research at CHU Sainte Justine's Pediatric Critical Care Unit (PICU) has revealed that traditional machine learning methods, such as semi-supervised label propagation and K-nearest neighbors, outperform Transformer-based models in artifact detection from PPG signals, mainly when data is limited. This study addresses the underutilization of abundant unlabeled data by employing self-supervised learning (SSL) to extract latent features from these data, followed by fine-tuning on labeled data. Our experiments demonstrate that SSL significantly enhances the Transformer model's ability to learn representations, improving its robustness in artifact classification tasks. Among various SSL techniques, including masking, contrastive learning, and DINO (self-distillation with no labels)-contrastive learning exhibited the most stable and superior performance in small PPG datasets. Further, we delve into optimizing contrastive loss functions, which are crucial for contrastive SSL. Inspired by InfoNCE, we introduce a novel contrastive loss function that facilitates smoother training and better convergence, thereby enhancing performance in artifact classification. In summary, this study establishes the efficacy of SSL in leveraging unlabeled data, particularly in enhancing the capabilities of the Transformer model. This approach holds promise for broader applications in PICU environments, where annotated data is often limited.
In this paper, we present a comparative analysis of various self-supervised Vision Transformers (ViTs), focusing on their local representative power. Inspired by large language models, we examine the abilities of ViTs to perform various computer vision tasks with little to no fine-tuning. We design evaluation framework to analyze the quality of local, i.e.\ patch-level, representations in the context of few-shot semantic segmentation, instance identification, object retrieval and tracking. We discover that contrastive learning based methods like DINO produce more universal patch representations that can be immediately applied for downstream tasks with no parameter tuning, compared to masked image modeling. The embeddings learned using the latter approach, e.g. in masked autoencoders, have high variance features that harm distance-based algorithms, such as k-NN, and do not contain useful information for most downstream tasks. Furthermore, we demonstrate that removing these high-variance features enhances k-NN for MAE, as well as for its recent extension Scale-MAE. Finally, we find an object instance retrieval setting where DINOv2, a model pretrained on two orders of magnitude more data, falls short of its less compute intensive counterpart DINO.
Medical data collected for diagnostic decisions are typically multimodal, providing comprehensive information on a subject. While computer-aided diagnosis systems can benefit from multimodal inputs, effectively fusing such data remains a challenging task and a key focus in medical research. In this paper, we propose a transformer-based framework, called Alifuse, for aligning and fusing multimodal medical data. Specifically, we convert medical images and both unstructured and structured clinical records into vision and language tokens, employing intramodal and intermodal attention mechanisms to learn unified representations of all imaging and non-imaging data for classification. Additionally, we integrate restoration modeling with contrastive learning frameworks, jointly learning the high-level semantic alignment between images and texts and the low-level understanding of one modality with the help of another. We apply Alifuse to classify Alzheimer's disease, achieving state-of-the-art performance on five public datasets and outperforming eight baselines.