Discriminative learning effectively predicts true object class for image classification. However, it often results in false positives for outliers, posing critical concerns in applications like autonomous driving and video surveillance systems. Previous attempts to address this challenge involved training image classifiers through contrastive learning using actual outlier data or synthesizing outliers for self-supervised learning. Furthermore, unsupervised generative modeling of inliers in pixel space has shown limited success for outlier detection. In this work, we introduce a quantile-based maximum likelihood objective for learning the inlier distribution to improve the outlier separation during inference. Our approach fits a normalizing flow to pre-trained discriminative features and detects the outliers according to the evaluated log-likelihood. The experimental evaluation demonstrates the effectiveness of our method as it surpasses the performance of the state-of-the-art unsupervised methods for outlier detection. The results are also competitive compared with a recent self-supervised approach for outlier detection. Our work allows to reduce dependency on well-sampled negative training data, which is especially important for domains like medical diagnostics or remote sensing.
Long scan time significantly hinders the widespread applications of three-dimensional multi-contrast cardiac magnetic resonance (3D-MC-CMR) imaging. This study aims to accelerate 3D-MC-CMR acquisition by a novel method based on score-based diffusion models with self-supervised learning. Specifically, we first establish a mapping between the undersampled k-space measurements and the MR images, utilizing a self-supervised Bayesian reconstruction network. Secondly, we develop a joint score-based diffusion model on 3D-MC-CMR images to capture their inherent distribution. The 3D-MC-CMR images are finally reconstructed using the conditioned Langenvin Markov chain Monte Carlo sampling. This approach enables accurate reconstruction without fully sampled training data. Its performance was tested on the dataset acquired by a 3D joint myocardial T1 and T1rho mapping sequence. The T1 and T1rho maps were estimated via a dictionary matching method from the reconstructed images. Experimental results show that the proposed method outperforms traditional compressed sensing and existing self-supervised deep learning MRI reconstruction methods. It also achieves high quality T1 and T1rho parametric maps close to the reference maps, even at a high acceleration rate of 14.
Self-supervised learning offers an efficient way of extracting rich representations from various types of unlabeled data while avoiding the cost of annotating large-scale datasets. This is achievable by designing a pretext task to form pseudo labels with respect to the modality and domain of the data. Given the evolving applications of online handwritten texts, in this study, we propose the novel Part of Stroke Masking (POSM) as a pretext task for pretraining models to extract informative representations from the online handwriting of individuals in English and Chinese languages, along with two suggested pipelines for fine-tuning the pretrained models. To evaluate the quality of the extracted representations, we use both intrinsic and extrinsic evaluation methods. The pretrained models are fine-tuned to achieve state-of-the-art results in tasks such as writer identification, gender classification, and handedness classification, also highlighting the superiority of utilizing the pretrained models over the models trained from scratch.
Dataset distillation methods have achieved remarkable success in distilling a large dataset into a small set of representative samples. However, they are not designed to produce a distilled dataset that can be effectively used for facilitating self-supervised pre-training. To this end, we propose a novel problem of distilling an unlabeled dataset into a set of small synthetic samples for efficient self-supervised learning (SSL). We first prove that a gradient of synthetic samples with respect to a SSL objective in naive bilevel optimization is \textit{biased} due to the randomness originating from data augmentations or masking. To address this issue, we propose to minimize the mean squared error (MSE) between a model's representations of the synthetic examples and their corresponding learnable target feature representations for the inner objective, which does not introduce any randomness. Our primary motivation is that the model obtained by the proposed inner optimization can mimic the \textit{self-supervised target model}. To achieve this, we also introduce the MSE between representations of the inner model and the self-supervised target model on the original full dataset for outer optimization. Lastly, assuming that a feature extractor is fixed, we only optimize a linear head on top of the feature extractor, which allows us to reduce the computational cost and obtain a closed-form solution of the head with kernel ridge regression. We empirically validate the effectiveness of our method on various applications involving transfer learning.
Self-supervised learning (SSL) for WiFi-based human activity recognition (HAR) holds great promise due to its ability to address the challenge of insufficient labeled data. However, directly transplanting SSL algorithms, especially contrastive learning, originally designed for other domains to CSI data, often fails to achieve the expected performance. We attribute this issue to the inappropriate alignment criteria, which disrupt the semantic distance consistency between the feature space and the input space. To address this challenge, we introduce \textbf{A}ntenna \textbf{R}esponse \textbf{C}onsistency (ARC) as a solution to define proper alignment criteria. ARC is designed to retain semantic information from the input space while introducing robustness to real-world noise. Moreover, we substantiate the effectiveness of ARC through a comprehensive set of experiments, demonstrating its capability to enhance the performance of self-supervised learning for WiFi-based HAR by achieving an increase of over 5\% in accuracy in most cases and achieving a best accuracy of 94.97\%.
Sickle Cell Disease (SCD) is a chronic genetic disorder characterized by recurrent acute painful episodes. Opioids are often used to manage these painful episodes; the extent of their use in managing pain in this disorder is an issue of debate. The risk of addiction and side effects of these opioid treatments can often lead to more pain episodes in the future. Hence, it is crucial to forecast future patient pain trajectories to help patients manage their SCD to improve their quality of life without compromising their treatment. It is challenging to obtain many pain records to design forecasting models since it is mainly recorded by patients' self-report. Therefore, it is expensive and painful (due to the need for patient compliance) to solve pain forecasting problems in a purely supervised manner. In light of this challenge, we propose to solve the pain forecasting problem using self-supervised learning methods. Also, clustering such time-series data is crucial for patient phenotyping, anticipating patients' prognoses by identifying "similar" patients, and designing treatment guidelines tailored to homogeneous patient subgroups. Hence, we propose a self-supervised learning approach for clustering time-series data, where each cluster comprises patients who share similar future pain profiles. Experiments on five years of real-world datasets show that our models achieve superior performance over state-of-the-art benchmarks and identify meaningful clusters that can be translated into actionable information for clinical decision-making.
Cross-domain sequential recommenders (CSRs) are gaining considerable research attention as they can capture user sequential preference by leveraging side information from multiple domains. However, these works typically follow an ideal setup, i.e., different domains obey similar data distribution, which ignores the bias brought by asymmetric interaction densities (a.k.a. the inter-domain density bias). Besides, the frequently adopted mechanism (e.g., the self-attention network) in sequence encoder only focuses on the interactions within a local view, which overlooks the global correlations between different training batches. To this end, we propose an External Attention-enhanced Graph Contrastive Learning framework, namely EA-GCL. Specifically, to remove the impact of the inter-domain density bias, an auxiliary Self-Supervised Learning (SSL) task is attached to the traditional graph encoder under a multi-task learning manner. To robustly capture users' behavioral patterns, we develop an external attention-based sequence encoder that contains an MLP-based memory-sharing structure. Unlike the self-attention mechanism, such a structure can effectively alleviate the bias interference from the batch-based training scheme. Extensive experiments on two real-world datasets demonstrate that EA-GCL outperforms several state-of-the-art baselines on CSR tasks. The source codes and relevant datasets are available at https://github.com/HoupingY/EA-GCL.
Speech and language models trained through self-supervised learning (SSL) demonstrate strong alignment with brain activity during speech and language perception. However, given their distinct training modalities, it remains unclear whether they correlate with the same neural aspects. We directly address this question by evaluating the brain prediction performance of two representative SSL models, Wav2Vec2.0 and GPT-2, designed for speech and language tasks. Our findings reveal that both models accurately predict speech responses in the auditory cortex, with a significant correlation between their brain predictions. Notably, shared speech contextual information between Wav2Vec2.0 and GPT-2 accounts for the majority of explained variance in brain activity, surpassing static semantic and lower-level acoustic-phonetic information. These results underscore the convergence of speech contextual representations in SSL models and their alignment with the neural network underlying speech perception, offering valuable insights into both SSL models and the neural basis of speech and language processing.
Beyond the great cognitive powers showcased by language models, it is crucial to scrutinize whether their reasoning capabilities stem from strong generalization or merely exposure to relevant data. As opposed to constructing increasingly complex logic, this paper probes into the boolean logic, the root capability of a logical reasoner. We find that any pre-trained language models even including large language models only behave like a random selector in the face of multi-nested boolean logic, a task that humans can handle with ease. To empower language models with this fundamental capability, this paper proposes a new self-supervised learning method \textit{Curriculum Logical Reasoning} (\textsc{Clr}), where we augment the training data with nested boolean logic chain step-by-step, and program the training from simpler logical patterns gradually to harder ones. This new training paradigm allows language models to effectively generalize to much harder and longer-hop logic, which can hardly be learned through naive training. Furthermore, we show that boolean logic is a great foundation for improving the subsequent general logical tasks.
In this paper we introduce a new synchronisation task, Gesture-Sync: determining if a person's gestures are correlated with their speech or not. In comparison to Lip-Sync, Gesture-Sync is far more challenging as there is a far looser relationship between the voice and body movement than there is between voice and lip motion. We introduce a dual-encoder model for this task, and compare a number of input representations including RGB frames, keypoint images, and keypoint vectors, assessing their performance and advantages. We show that the model can be trained using self-supervised learning alone, and evaluate its performance on the LRS3 dataset. Finally, we demonstrate applications of Gesture-Sync for audio-visual synchronisation, and in determining who is the speaker in a crowd, without seeing their faces. The code, datasets and pre-trained models can be found at: \url{https://www.robots.ox.ac.uk/~vgg/research/gestsync}.
This paper presents our systems (denoted as T13) for the singing voice conversion challenge (SVCC) 2023. For both in-domain and cross-domain English singing voice conversion (SVC) tasks (Task 1 and Task 2), we adopt a recognition-synthesis approach with self-supervised learning-based representation. To achieve data-efficient SVC with a limited amount of target singer/speaker's data (150 to 160 utterances for SVCC 2023), we first train a diffusion-based any-to-any voice conversion model using publicly available large-scale 750 hours of speech and singing data. Then, we finetune the model for each target singer/speaker of Task 1 and Task 2. Large-scale listening tests conducted by SVCC 2023 show that our T13 system achieves competitive naturalness and speaker similarity for the harder cross-domain SVC (Task 2), which implies the generalization ability of our proposed method. Our objective evaluation results show that using large datasets is particularly beneficial for cross-domain SVC.
Incorporating heterogeneous representations from different architectures has facilitated various vision tasks, e.g., some hybrid networks combine transformers and convolutions. However, complementarity between such heterogeneous architectures has not been well exploited in self-supervised learning. Thus, we propose Heterogeneous Self-Supervised Learning (HSSL), which enforces a base model to learn from an auxiliary head whose architecture is heterogeneous from the base model. In this process, HSSL endows the base model with new characteristics in a representation learning way without structural changes. To comprehensively understand the HSSL, we conduct experiments on various heterogeneous pairs containing a base model and an auxiliary head. We discover that the representation quality of the base model moves up as their architecture discrepancy grows. This observation motivates us to propose a search strategy that quickly determines the most suitable auxiliary head for a specific base model to learn and several simple but effective methods to enlarge the model discrepancy. The HSSL is compatible with various self-supervised methods, achieving superior performances on various downstream tasks, including image classification, semantic segmentation, instance segmentation, and object detection. Our source code will be made publicly available.
The ability to detect objects in images at varying scales has played a pivotal role in the design of modern object detectors. Despite considerable progress in removing hand-crafted components and simplifying the architecture with transformers, multi-scale feature maps and/or pyramid design remain a key factor for their empirical success. In this paper, we show that this reliance on either feature pyramids or an hierarchical backbone is unnecessary and a transformer-based detector with scale-aware attention enables the plain detector `SimPLR' whose backbone and detection head are both non-hierarchical and operate on single-scale features. We find through our experiments that SimPLR with scale-aware attention is plain and simple, yet competitive with multi-scale vision transformer alternatives. Compared to the multi-scale and single-scale state-of-the-art, our model scales much better with bigger capacity (self-supervised) models and more pre-training data, allowing us to report a consistently better accuracy and faster runtime for object detection, instance segmentation as well as panoptic segmentation. Code will be released.
Spatio-temporal predictive learning plays a crucial role in self-supervised learning, with wide-ranging applications across a diverse range of fields. Previous approaches for temporal modeling fall into two categories: recurrent-based and recurrent-free methods. The former, while meticulously processing frames one by one, neglect short-term spatio-temporal information redundancies, leading to inefficiencies. The latter naively stack frames sequentially, overlooking the inherent temporal dependencies. In this paper, we re-examine the two dominant temporal modeling approaches within the realm of spatio-temporal predictive learning, offering a unified perspective. Building upon this analysis, we introduce USTEP (Unified Spatio-TEmporal Predictive learning), an innovative framework that reconciles the recurrent-based and recurrent-free methods by integrating both micro-temporal and macro-temporal scales. Extensive experiments on a wide range of spatio-temporal predictive learning demonstrate that USTEP achieves significant improvements over existing temporal modeling approaches, thereby establishing it as a robust solution for a wide range of spatio-temporal applications.
Image-based depth estimation has gained significant attention in recent research on computer vision for autonomous vehicles in intelligent transportation systems. This focus stems from its cost-effectiveness and wide range of potential applications. Unlike binocular depth estimation methods that require two fixed cameras, monocular depth estimation methods only rely on a single camera, making them highly versatile. While state-of-the-art approaches for this task leverage self-supervised learning of deep neural networks in conjunction with tasks like pose estimation and semantic segmentation, none of them have explored the combination of federated learning and self-supervision to train models using unlabeled and private data captured by autonomous vehicles. The utilization of federated learning offers notable benefits, including enhanced privacy protection, reduced network consumption, and improved resilience to connectivity issues. To address this gap, we propose FedSCDepth, a novel method that combines federated learning and deep self-supervision to enable the learning of monocular depth estimators with comparable effectiveness and superior efficiency compared to the current state-of-the-art methods. Our evaluation experiments conducted on Eigen's Split of the KITTI dataset demonstrate that our proposed method achieves near state-of-the-art performance, with a test loss below 0.13 and requiring, on average, only 1.5k training steps and up to 0.415 GB of weight data transfer per autonomous vehicle on each round.
When delineating lesions from medical images, a human expert can always keep in mind the anatomical structure behind the voxels. However, although high-quality (though not perfect) anatomical information can be retrieved from computed tomography (CT) scans with modern deep learning algorithms, it is still an open problem how these automatically generated organ masks can assist in addressing challenging lesion segmentation tasks, such as the segmentation of colorectal cancer (CRC). In this paper, we develop a novel Anatomy-Guided segmentation framework to exploit the auto-generated organ masks to aid CRC segmentation from CT, namely AG-CRC. First, we obtain multi-organ segmentation (MOS) masks with existing MOS models (e.g., TotalSegmentor) and further derive a more robust organ of interest (OOI) mask that may cover most of the colon-rectum and CRC voxels. Then, we propose an anatomy-guided training patch sampling strategy by optimizing a heuristic gain function that considers both the proximity of important regions (e.g., the tumor or organs of interest) and sample diversity. Third, we design a novel self-supervised learning scheme inspired by the topology of tubular organs like the colon to boost the model performance further. Finally, we employ a masked loss scheme to guide the model to focus solely on the essential learning region. We extensively evaluate the proposed method on two CRC segmentation datasets, where substantial performance improvement (5% to 9% in Dice) is achieved over current state-of-the-art medical image segmentation models, and the ablation studies further evidence the efficacy of every proposed component.
Addressing the challenge of organizing scattered items in domestic spaces is complicated by the diversity and subjective nature of tidiness. Just as the complexity of human language allows for multiple expressions of the same idea, household tidiness preferences and organizational patterns vary widely, so presetting object locations would limit the adaptability to new objects and environments. Inspired by advancements in natural language processing (NLP), this paper introduces a self-supervised learning framework that allows robots to understand and replicate the concept of tidiness from demonstrations of well-organized layouts, akin to using conversational datasets to train Large Language Models(LLM). We leverage a transformer neural network to predict the placement of subsequent objects. We demonstrate a ``knolling'' system with a robotic arm and an RGB camera to organize items of varying sizes and quantities on a table. Our method not only trains a generalizable concept of tidiness, enabling the model to provide diverse solutions and adapt to different numbers of objects, but it can also incorporate human preferences to generate customized tidy tables without explicit target positions for each object.
Pre-training has been an important ingredient in developing strong monocular depth estimation models in recent years. For instance, self-supervised learning (SSL) is particularly effective by alleviating the need for large datasets with dense ground-truth depth maps. However, despite these improvements, our study reveals that the later layers of the SOTA SSL method are actually suboptimal. By examining the layer-wise representations, we demonstrate significant changes in these later layers during fine-tuning, indicating the ineffectiveness of their pre-trained features for depth estimation. To address these limitations, we propose MeSa, a comprehensive framework that leverages the complementary strengths of masked, geometric, and supervised pre-training. Hence, MeSa benefits from not only general-purpose representations learnt via masked pre training but also specialized depth-specific features acquired via geometric and supervised pre-training. Our CKA layer-wise analysis confirms that our pre-training strategy indeed produces improved representations for the later layers, overcoming the drawbacks of the SOTA SSL method. Furthermore, via experiments on the NYUv2 and IBims-1 datasets, we demonstrate that these enhanced representations translate to performance improvements in both the in-distribution and out-of-distribution settings. We also investigate the influence of the pre-training dataset and demonstrate the efficacy of pre-training on LSUN, which yields significantly better pre-trained representations. Overall, our approach surpasses the masked pre-training SSL method by a substantial margin of 17.1% on the RMSE. Moreover, even without utilizing any recently proposed techniques, MeSa also outperforms the most recent methods and establishes a new state-of-the-art for monocular depth estimation on the challenging NYUv2 dataset.
Many Self-Supervised Learning (SSL) methods aim for model invariance to different image augmentations known as views. To achieve this invariance, conventional approaches make use of random sampling operations within the image augmentation pipeline. We hypothesize that the efficacy of pretraining pipelines based on conventional random view sampling can be enhanced by explicitly selecting views that benefit the learning progress. A simple, yet effective approach is to select hard views that yield a higher loss. In this paper, we present Hard View Pretraining (HVP), a learning-free strategy that builds upon this hypothesis and extends random view generation. HVP exposes the model to harder, more challenging samples during SSL pretraining, which enhances downstream performance. It encompasses the following iterative steps: 1) randomly sample multiple views and forward each view through the pretrained model, 2) create pairs of two views and compute their loss, 3) adversarially select the pair yielding the highest loss depending on the current model state, and 4) run the backward pass with the selected pair. As a result, HVP achieves linear evaluation accuracy improvements of 1% on average on ImageNet for both 100 and 300 epoch pretraining and similar improvements on transfer tasks across DINO, SimSiam, iBOT, and SimCLR.
A foundation model is a machine learning model trained on a large and diverse set of data, typically using self-supervised learning-based pre-training techniques, that can be adapted to various downstream tasks. However, current research on time series pre-training has mostly focused on models pre-trained solely on data from a single domain, resulting in a lack of knowledge about other types of time series. However, current research on time series pre-training has predominantly focused on models trained exclusively on data from a single domain. As a result, these models possess domain-specific knowledge that may not be easily transferable to time series from other domains. In this paper, we aim to develop an effective time series foundation model by leveraging unlabeled samples from multiple domains. To achieve this, we repurposed the publicly available UCR Archive and evaluated four existing self-supervised learning-based pre-training methods, along with a novel method, on the datasets. We tested these methods using four popular neural network architectures for time series to understand how the pre-training methods interact with different network designs. Our experimental results show that pre-training improves downstream classification tasks by enhancing the convergence of the fine-tuning process. Furthermore, we found that the proposed pre-training method, when combined with the Transformer model, outperforms the alternatives.