Self-supervised foundation models have recently been successfully extended to encode three-dimensional (3D) computed tomography (CT) images, with excellent performance across several downstream tasks, such as intracranial hemorrhage detection and lung cancer risk forecasting. However, as self-supervised models learn from complex data distributions, questions arise concerning whether these embeddings capture demographic information, such as age, sex, or race. Using the National Lung Screening Trial (NLST) dataset, which contains 3D CT images and demographic data, we evaluated a range of classifiers: softmax regression, linear regression, linear support vector machine, random forest, and decision tree, to predict sex, race, and age of the patients in the images. Our results indicate that the embeddings effectively encoded age and sex information, with a linear regression model achieving a root mean square error (RMSE) of 3.8 years for age prediction and a softmax regression model attaining an AUC of 0.998 for sex classification. Race prediction was less effective, with an AUC of 0.878. These findings suggest a detailed exploration into the information encoded in self-supervised learning frameworks is needed to help ensure fair, responsible, and patient privacy-protected healthcare AI.
Several self-supervised learning (SSL) approaches have shown that redundancy reduction in the feature embedding space is an effective tool for representation learning. However, these methods consider a narrow notion of redundancy, focusing on pairwise correlations between features. To address this limitation, we formalize the notion of embedding space redundancy and introduce redundancy measures that capture more complex, higher-order dependencies. We mathematically analyze the relationships between these metrics, and empirically measure these redundancies in the embedding spaces of common SSL methods. Based on our findings, we propose Self Supervised Learning with Predictability Minimization (SSLPM) as a method for reducing redundancy in the embedding space. SSLPM combines an encoder network with a predictor engaging in a competitive game of reducing and exploiting dependencies respectively. We demonstrate that SSLPM is competitive with state-of-the-art methods and find that the best performing SSL methods exhibit low embedding space redundancy, suggesting that even methods without explicit redundancy reduction mechanisms perform redundancy reduction implicitly.
This paper investigates the impact of sampling and pretraining using datasets with different image characteristics on the performance of self-supervised learning (SSL) models for object classification. To do this, we sample two apartment datasets from the Omnidata platform based on modality, luminosity, image size, and camera field of view and use them to pretrain a SimCLR model. The encodings generated from the pretrained model are then transferred to a supervised Resnet-50 model for object classification. Through A/B testing, we find that depth pretrained models are more effective on low resolution images, while RGB pretrained models perform better on higher resolution images. We also discover that increasing the luminosity of training images can improve the performance of models on low resolution images without negatively affecting their performance on higher resolution images.
We propose a method to improve the generalization ability of skin lesion classification models by combining self-supervised learning (SSL), unsupervised domain adaptation (UDA), and active domain adaptation (ADA). The main steps of the approach include selection of a SSL pretrained model on natural image datasets, subsequent SSL retraining on all available skin lesion datasets, finetuning of the model on source domain data with labels, application of UDA methods on target domain data, and lastly, implementation of ADA methods. The efficacy of the proposed approach is assessed across ten skin lesion datasets of domains, demonstrating its potential for enhancing the performance of skin lesion classification models. This approach holds promise for facilitating the widespread adoption of medical imaging models in clinical settings, thereby amplifying their impact.
Self-supervised learning (SSL) methods have emerged as strong visual representation learners by training an image encoder to maximize similarity between features of different views of the same image. To perform this view-invariance task, current SSL algorithms rely on hand-crafted augmentations such as random cropping and color jittering to create multiple views of an image. Recently, generative diffusion models have been shown to improve SSL by providing a wider range of data augmentations. However, these diffusion models require pre-training on large-scale image-text datasets, which might not be available for many specialized domains like histopathology. In this work, we introduce Gen-SIS, a diffusion-based augmentation technique trained exclusively on unlabeled image data, eliminating any reliance on external sources of supervision such as text captions. We first train an initial SSL encoder on a dataset using only hand-crafted augmentations. We then train a diffusion model conditioned on embeddings from that SSL encoder. Following training, given an embedding of the source image, this diffusion model can synthesize its diverse views. We show that these `self-augmentations', i.e. generative augmentations based on the vanilla SSL encoder embeddings, facilitate the training of a stronger SSL encoder. Furthermore, based on the ability to interpolate between images in the encoder latent space, we introduce the novel pretext task of disentangling the two source images of an interpolated synthetic image. We validate Gen-SIS's effectiveness by demonstrating performance improvements across various downstream tasks in both natural images, which are generally object-centric, as well as digital histopathology images, which are typically context-based.
Data augmentation is widely used to enhance generalization in visual classification tasks. However, traditional methods struggle when source and target domains differ, as in domain adaptation, due to their inability to address domain gaps. This paper introduces GenMix, a generalizable prompt-guided generative data augmentation approach that enhances both in-domain and cross-domain image classification. Our technique leverages image editing to generate augmented images based on custom conditional prompts, designed specifically for each problem type. By blending portions of the input image with its edited generative counterpart and incorporating fractal patterns, our approach mitigates unrealistic images and label ambiguity, improving the performance and adversarial robustness of the resulting models. Efficacy of our method is established with extensive experiments on eight public datasets for general and fine-grained classification, in both in-domain and cross-domain settings. Additionally, we demonstrate performance improvements for self-supervised learning, learning with data scarcity, and adversarial robustness. As compared to the existing state-of-the-art methods, our technique achieves stronger performance across the board.
Self Supervised learning (SSL) has demonstrated its effectiveness in feature learning from unlabeled data. Regarding this success, there have been some arguments on the role that mutual information plays within the SSL framework. Some works argued for increasing mutual information between representation of augmented views. Others suggest decreasing mutual information between them, while increasing task-relevant information. We ponder upon this debate and propose to revisit the core idea of SSL within the framework of partial information decomposition (PID). Thus, with SSL under PID we propose to replace traditional mutual information with the more general concept of joint mutual information to resolve the argument. Our investigation on instantiation of SSL within the PID framework leads to upgrading the existing pipelines by considering the components of the PID in the SSL models for improved representation learning. Accordingly we propose a general pipeline that can be applied to improve existing baselines. Our pipeline focuses on extracting the unique information component under the PID to build upon lower level supervision for generic feature learning and on developing higher-level supervisory signals for task-related feature learning. In essence, this could be interpreted as a joint utilization of local and global clustering. Experiments on four baselines and four datasets show the effectiveness and generality of our approach in improving existing SSL frameworks.
The success of self-supervised learning (SSL) has been the focus of multiple recent theoretical and empirical studies, including the role of data augmentation (in feature decoupling) as well as complete and dimensional representation collapse. While complete collapse is well-studied and addressed, dimensional collapse has only gain attention and addressed in recent years mostly using variants of redundancy reduction (aka whitening) techniques. In this paper, we further explore a complementary approach to whitening via feature decoupling for improved representation learning while avoiding representation collapse. In particular, we perform feature decoupling by early promotion of useful features via careful feature coloring. The coloring technique is developed based on a Bayesian prior of the augmented data, which is inherently encoded for feature decoupling. We show that our proposed framework is complementary to the state-of-the-art techniques, while outperforming both contrastive and recent non-contrastive methods. We also study the different effects of coloring approach to formulate it as a general complementary technique along with other baselines.
Test-time adaptation (TTA) has emerged as a promising solution to tackle the continual domain shift in machine learning by allowing model parameters to change at test time, via self-supervised learning on unlabeled testing data. At the same time, it unfortunately opens the door to unforeseen vulnerabilities for degradation over time. Through a simple theoretical continual TTA model, we successfully identify a risk in the sampling process of testing data that could easily degrade the performance of a continual TTA model. We name this risk as Reusing of Incorrect Prediction (RIP) that TTA attackers can employ or as a result of the unintended query from general TTA users. The risk posed by RIP is also highly realistic, as it does not require prior knowledge of model parameters or modification of testing samples. This simple requirement makes RIP as the first black-box TTA attack algorithm that stands out from existing white-box attempts. We extensively benchmark the performance of the most recent continual TTA approaches when facing the RIP attack, providing insights on its success, and laying out potential roadmaps that could enhance the resilience of future continual TTA systems.
Representations learned by self-supervised approaches are generally considered to possess sufficient generalizability and discriminability. However, we disclose a nontrivial mutual-exclusion relationship between these critical representation properties through an exploratory demonstration on self-supervised learning. State-of-the-art self-supervised methods tend to enhance either generalizability or discriminability but not both simultaneously. Thus, learning representations jointly possessing strong generalizability and discriminability presents a specific challenge for self-supervised learning. To this end, we revisit the learning paradigm of self-supervised learning from the perspective of evolutionary game theory (EGT) and outline the theoretical roadmap to achieve a desired trade-off between these representation properties. EGT performs well in analyzing the trade-off point in a two-player game by utilizing dynamic system modeling. However, the EGT analysis requires sufficient annotated data, which contradicts the principle of self-supervised learning, i.e., the EGT analysis cannot be conducted without the annotations of the specific target domain for self-supervised learning. Thus, to enhance the methodological generalization, we propose a novel self-supervised learning method that leverages advancements in reinforcement learning to jointly benefit from the general guidance of EGT and sequentially optimize the model to chase the consistent improvement of generalizability and discriminability for specific target domains during pre-training. Theoretically, we establish that the proposed method tightens the generalization error upper bound of self-supervised learning. Empirically, our method achieves state-of-the-art performance on various benchmarks.
Vision-Language Models (VLMs) trained with contrastive loss have achieved significant advancements in various vision and language tasks. However, the global nature of the contrastive loss makes VLMs focus predominantly on foreground objects, neglecting other crucial information in the image, which limits their effectiveness in downstream tasks. To address these challenges, we propose COSMOS: CrOSs-MOdality Self-distillation for vision-language pre-training that integrates a novel text-cropping strategy and cross-attention module into a self-supervised learning framework. We create global and local views of images and texts (i.e., multi-modal augmentations), which are essential for self-distillation in VLMs. We further introduce a cross-attention module, enabling COSMOS to learn comprehensive cross-modal representations optimized via a cross-modality self-distillation loss. COSMOS consistently outperforms previous strong baselines on various zero-shot downstream tasks, including retrieval, classification, and semantic segmentation. Additionally, it surpasses CLIP-based models trained on larger datasets in visual perception and contextual understanding tasks. Code is available at https://github.com/ExplainableML/cosmos.
This paper introduces SmartBSP, an advanced self-supervised learning framework for real-time path planning and obstacle avoidance in autonomous robotics navigating through complex environments. The proposed system integrates Proximal Policy Optimization (PPO) with Convolutional Neural Networks (CNN) and Actor-Critic architecture to process limited LIDAR inputs and compute spatial decision-making probabilities. The robot's perceptual field is discretized into a grid format, which the CNN analyzes to produce a spatial probability distribution. During the training process a nuanced cost function is minimized that accounts for path curvature, endpoint proximity, and obstacle avoidance. Simulations results in different scenarios validate the algorithm's resilience and adaptability across diverse operational scenarios. Subsequently, Real-time experiments, employing the Robot Operating System (ROS), were carried out to assess the efficacy of the proposed algorithm.
One-shot voice conversion (VC) aims to alter the timbre of speech from a source speaker to match that of a target speaker using just a single reference speech from the target, while preserving the semantic content of the original source speech. Despite advancements in one-shot VC, its effectiveness decreases in real-world scenarios where reference speeches, often sourced from the internet, contain various disturbances like background noise. To address this issue, we introduce Noro, a Noise Robust One-shot VC system. Noro features innovative components tailored for VC using noisy reference speeches, including a dual-branch reference encoding module and a noise-agnostic contrastive speaker loss. Experimental results demonstrate that Noro outperforms our baseline system in both clean and noisy scenarios, highlighting its efficacy for real-world applications. Additionally, we investigate the hidden speaker representation capabilities of our baseline system by repurposing its reference encoder as a speaker encoder. The results shows that it is competitive with several advanced self-supervised learning models for speaker representation under the SUPERB settings, highlighting the potential for advancing speaker representation learning through one-shot VC task.
Pre-training on large-scale unlabeled datasets contribute to the model achieving powerful performance on 3D vision tasks, especially when annotations are limited. However, existing rendering-based self-supervised frameworks are computationally demanding and memory-intensive during pre-training due to the inherent nature of volume rendering. In this paper, we propose an efficient framework named GS$^3$ to learn point cloud representation, which seamlessly integrates fast 3D Gaussian Splatting into the rendering-based framework. The core idea behind our framework is to pre-train the point cloud encoder by comparing rendered RGB images with real RGB images, as only Gaussian points enriched with learned rich geometric and appearance information can produce high-quality renderings. Specifically, we back-project the input RGB-D images into 3D space and use a point cloud encoder to extract point-wise features. Then, we predict 3D Gaussian points of the scene from the learned point cloud features and uses a tile-based rasterizer for image rendering. Finally, the pre-trained point cloud encoder can be fine-tuned to adapt to various downstream 3D tasks, including high-level perception tasks such as 3D segmentation and detection, as well as low-level tasks such as 3D scene reconstruction. Extensive experiments on downstream tasks demonstrate the strong transferability of the pre-trained point cloud encoder and the effectiveness of our self-supervised learning framework. In addition, our GS$^3$ framework is highly efficient, achieving approximately 9$\times$ pre-training speedup and less than 0.25$\times$ memory cost compared to the previous rendering-based framework Ponder.
In the era of foundation models, CLIP has emerged as a powerful tool for aligning text and visual modalities into a common embedding space. However, the alignment objective used to train CLIP often results in subpar visual features for fine-grained tasks. In contrast, SSL-pretrained models like DINO excel at extracting rich visual features due to their specialized training paradigm. Yet, these SSL models require an additional supervised linear probing step, which relies on fully labeled data which is often expensive and difficult to obtain at scale. In this paper, we propose a label-free prompt-tuning method that leverages the rich visual features of self-supervised learning models (DINO) and the broad textual knowledge of large language models (LLMs) to largely enhance CLIP-based image classification performance using unlabeled images. Our approach unfolds in three key steps: (1) We generate robust textual feature embeddings that more accurately represent object classes by leveraging class-specific descriptions from LLMs, enabling more effective zero-shot classification compared to CLIP's default name-specific prompts. (2) These textual embeddings are then used to produce pseudo-labels to train an alignment module that integrates the complementary strengths of LLM description-based textual embeddings and DINO's visual features. (3) Finally, we prompt-tune CLIP's vision encoder through DINO-assisted supervision using the trained alignment module. This three-step process allows us to harness the best of visual and textual foundation models, resulting in a powerful and efficient approach that surpasses state-of-the-art label-free classification methods. Notably, our framework, NoLA (No Labels Attached), achieves an average absolute gain of 3.6% over the state-of-the-art LaFter across 11 diverse image classification datasets.
We present RelCon, a novel self-supervised Relative Contrastive learning approach for training a motion foundation model from wearable accelerometry sensors. First, a learnable distance measure is trained to capture motif similarity and domain-specific semantic information such as rotation invariance. Then, the learned distance provides a measurement of semantic similarity between a pair of accelerometry time-series, which we use to train our foundation model to model relative relationships across time and across subjects. The foundation model is trained on 1 billion segments from 87,376 participants, and achieves state-of-the-art performance across multiple downstream tasks, including human activity recognition and gait metric regression. To our knowledge, we are the first to show the generalizability of a foundation model with motion data from wearables across distinct evaluation tasks.
The field of computational pathology has been transformed with recent advances in foundation models that encode histopathology region-of-interests (ROIs) into versatile and transferable feature representations via self-supervised learning (SSL). However, translating these advancements to address complex clinical challenges at the patient and slide level remains constrained by limited clinical data in disease-specific cohorts, especially for rare clinical conditions. We propose TITAN, a multimodal whole slide foundation model pretrained using 335,645 WSIs via visual self-supervised learning and vision-language alignment with corresponding pathology reports and 423,122 synthetic captions generated from a multimodal generative AI copilot for pathology. Without any finetuning or requiring clinical labels, TITAN can extract general-purpose slide representations and generate pathology reports that generalize to resource-limited clinical scenarios such as rare disease retrieval and cancer prognosis. We evaluate TITAN on diverse clinical tasks and find that TITAN outperforms both ROI and slide foundation models across machine learning settings such as linear probing, few-shot and zero-shot classification, rare cancer retrieval and cross-modal retrieval, and pathology report generation.
Self-supervised learning (SSL) has achieved great success in speech-related tasks. While Transformer and Conformer architectures have dominated SSL backbones, encoders like Zipformer, which excel in automatic speech recognition (ASR), remain unexplored in SSL. Concurrently, inefficiencies in data processing within existing SSL training frameworks, such as fairseq, pose challenges in managing the growing volumes of training data. To address these issues, we propose k2SSL, an open-source framework that offers faster, more memory-efficient, and better-performing self-supervised speech representation learning, focusing on downstream ASR tasks. The optimized HuBERT and proposed Zipformer-based SSL systems exhibit substantial reductions in both training time and memory usage during SSL training. Experiments on LibriSpeech demonstrate that Zipformer Base significantly outperforms HuBERT and WavLM, achieving up to a 34.8% relative WER reduction compared to HuBERT Base after fine-tuning, along with a 3.5x pre-training speedup in GPU hours. When scaled to 60k hours of LibriLight data, Zipformer Large exhibits remarkable efficiency, matching HuBERT Large's performance while requiring only 5/8 pre-training steps.
In recent years, graph representation learning has undergone a paradigm shift, driven by the emergence and proliferation of graph neural networks (GNNs) and their heterogeneous counterparts. Heterogeneous GNNs have shown remarkable success in extracting low-dimensional embeddings from complex graphs that encompass diverse entity types and relationships. While meta-path-based techniques have long been recognized for their ability to capture semantic affinities among nodes, their dependence on manual specification poses a significant limitation. In contrast, matrix-focused methods accelerate processing by utilizing structural cues but often overlook contextual richness. In this paper, we challenge the current paradigm by introducing ontology as a fundamental semantic primitive within complex graphs. Our goal is to integrate the strengths of both matrix-centric and meta-path-based approaches into a unified framework. We propose perturbation Ontology-based Graph Attention Networks (POGAT), a novel methodology that combines ontology subgraphs with an advanced self-supervised learning paradigm to achieve a deep contextual understanding. The core innovation of POGAT lies in our enhanced homogeneous perturbing scheme designed to generate rigorous negative samples, encouraging the model to explore minimal contextual features more thoroughly. Through extensive empirical evaluations, we demonstrate that POGAT significantly outperforms state-of-the-art baselines, achieving a groundbreaking improvement of up to 10.78\% in F1-score for the critical task of link prediction and 12.01\% in Micro-F1 for the critical task of node classification.
The utilization of speech Self-Supervised Learning (SSL) models achieves impressive performance on Automatic Speech Recognition (ASR). However, in low-resource language ASR, they encounter the domain mismatch problem between pre-trained and low-resource languages. Typical solutions like fine-tuning the SSL model suffer from high computation costs while using frozen SSL models as feature extractors comes with poor performance. To handle these issues, we extend a conventional efficient fine-tuning scheme based on the adapter. We add an extra intermediate adaptation to warm up the adapter and downstream model initialization. Remarkably, we update only 1-5% of the total model parameters to achieve the adaptation. Experimental results on the ML-SUPERB dataset show that our solution outperforms conventional efficient fine-tuning. It achieves up to a 28% relative improvement in the Character/Phoneme error rate when adapting to unseen languages.