Current emotional text-to-speech systems face challenges in conveying the full spectrum of human emotions, largely due to the inherent complexity of human emotions and the limited range of emotional labels in existing speech datasets. To address these limitations, this paper introduces a TTS framework that provides flexible user control over three emotional dimensions - pleasure, arousal, and dominance - enabling the synthesis of a diverse array of emotional styles. The framework leverages an emotional dimension predictor, trained soley on categorical labels from speech data and grounded in earlier psychological research, which is seamlessly integrated into a language model-based TTS system. Experimental results demonstrates that the proposed framework effectively learns emotional styles from expressive speech, eliminating the need for explicit emotion labels during TTS training, while enhancing the naturalness and diversity of synthesized emotional speech.
Face Forgery Detection (FFD), or Deepfake detection, aims to determine whether a digital face is real or fake. Due to different face synthesis algorithms with diverse forgery patterns, FFD models often overfit specific patterns in training datasets, resulting in poor generalization to other unseen forgeries. This severe challenge requires FFD models to possess strong capabilities in representing complex facial features and extracting subtle forgery cues. Although previous FFD models directly employ existing backbones to represent and extract facial forgery cues, the critical role of backbones is often overlooked, particularly as their knowledge and capabilities are insufficient to address FFD challenges, inevitably limiting generalization. Therefore, it is essential to integrate the backbone pre-training configurations and seek practical solutions by revisiting the complete FFD workflow, from backbone pre-training and fine-tuning to inference of discriminant results. Specifically, we analyze the crucial contributions of backbones with different configurations in FFD task and propose leveraging the ViT network with self-supervised learning on real-face datasets to pre-train a backbone, equipping it with superior facial representation capabilities. We then build a competitive backbone fine-tuning framework that strengthens the backbone's ability to extract diverse forgery cues within a competitive learning mechanism. Moreover, we devise a threshold optimization mechanism that utilizes prediction confidence to improve the inference reliability. Comprehensive experiments demonstrate that our FFD model with the elaborate backbone achieves excellent performance in FFD and extra face-related tasks, i.e., presentation attack detection. Code and models are available at https://github.com/zhenglab/FFDBackbone.
Utilizing Self-Supervised Learning (SSL) models for Speech Emotion Recognition (SER) has proven effective, yet limited research has explored cross-lingual scenarios. This study presents a comparative analysis between human performance and SSL models, beginning with a layer-wise analysis and an exploration of parameter-efficient fine-tuning strategies in monolingual, cross-lingual, and transfer learning contexts. We further compare the SER ability of models and humans at both utterance- and segment-levels. Additionally, we investigate the impact of dialect on cross-lingual SER through human evaluation. Our findings reveal that models, with appropriate knowledge transfer, can adapt to the target language and achieve performance comparable to native speakers. We also demonstrate the significant effect of dialect on SER for individuals without prior linguistic and paralinguistic background. Moreover, both humans and models exhibit distinct behaviors across different emotions. These results offer new insights into the cross-lingual SER capabilities of SSL models, underscoring both their similarities to and differences from human emotion perception.
Deep learning models have shown immense promise in computational pathology (CPath) tasks, but their performance often suffers when applied to unseen data due to domain shifts. Addressing this requires domain generalization (DG) algorithms. However, a systematic evaluation of DG algorithms in the CPath context is lacking. This study aims to benchmark the effectiveness of 30 DG algorithms on 3 CPath tasks of varying difficulty through 7,560 cross-validation runs. We evaluate these algorithms using a unified and robust platform, incorporating modality-specific techniques and recent advances like pretrained foundation models. Our extensive cross-validation experiments provide insights into the relative performance of various DG strategies. We observe that self-supervised learning and stain augmentation consistently outperform other methods, highlighting the potential of pretrained models and data augmentation. Furthermore, we introduce a new pan-cancer tumor detection dataset (HISTOPANTUM) as a benchmark for future research. This study offers valuable guidance to researchers in selecting appropriate DG approaches for CPath tasks.
Large language models (LLMs) have revolutionized how we interact with technology, but their personalization to individual user preferences remains a significant challenge, particularly in on-device applications. Traditional methods often depend heavily on labeled datasets and can be resource-intensive. To address these issues, we present Adaptive Self-Supervised Learning Strategies (ASLS), which utilizes self-supervised learning techniques to personalize LLMs dynamically. The framework comprises a user profiling layer for collecting interaction data and a neural adaptation layer for real-time model fine-tuning. This innovative approach enables continuous learning from user feedback, allowing the model to generate responses that align closely with user-specific contexts. The adaptive mechanisms of ASLS minimize computational demands and enhance personalization efficiency. Experimental results across various user scenarios illustrate the superior performance of ASLS in boosting user engagement and satisfaction, highlighting its potential to redefine LLMs as highly responsive and context-aware systems on-device.
Irregular time series, where data points are recorded at uneven intervals, are prevalent in healthcare settings, such as emergency wards where vital signs and laboratory results are captured at varying times. This variability, which reflects critical fluctuations in patient health, is essential for informed clinical decision-making. Existing self-supervised learning research on irregular time series often relies on generic pretext tasks like forecasting, which may not fully utilise the signal provided by irregular time series. There is a significant need for specialised pretext tasks designed for the characteristics of irregular time series to enhance model performance and robustness, especially in scenarios with limited data availability. This paper proposes a novel pretraining framework, EMIT, an event-based masking for irregular time series. EMIT focuses on masking-based reconstruction in the latent space, selecting masking points based on the rate of change in the data. This method preserves the natural variability and timing of measurements while enhancing the model's ability to process irregular intervals without losing essential information. Extensive experiments on the MIMIC-III and PhysioNet Challenge datasets demonstrate the superior performance of our event-based masking strategy. The code has been released at https://github.com/hrishi-ds/EMIT.
<div><p>In numerous inverse problems, state-of-the-art solving strategies involve training neural networks from ground truth and associated measurement datasets that, however, may be expensive or impossible to collect. Recently, self-supervised learning techniques have emerged, with the major advantage of no longer requiring ground truth data. Most theoretical and experimental results on self-supervised learning focus on linear inverse problems. The present work aims to study self-supervised learning for the non-linear inverse problem of recovering audio signals from clipped measurements. An equivariance-based selfsupervised loss is proposed and studied. Performance is assessed on simulated clipped measurements with controlled and varied levels of clipping, and further reported on standard real music signals. We show that the performance of the proposed equivariance-based self-supervised declipping strategy compares favorably to fully supervised learning while only requiring clipped measurements alone for training.</p></div>
We propose PseudoNeg-MAE, a novel self-supervised learning framework that enhances global feature representation of point cloud masked autoencoder by making them both discriminative and sensitive to transformations. Traditional contrastive learning methods focus on achieving invariance, discarding transformation-specific information. Recent approaches incorporate transformation sensitivity by explicitly modeling relationships between original and transformed inputs. However, they report an invariant-collapse phenomenon, where the predictor degenerates into identity mappings, resulting in latent representations that have limited variation across transformations. We propose a novel loss that explicitly penalizes invariant collapse, enabling the network to capture richer transformation cues while preserving discriminative representations. PseudoNeg-MAE uses a parametric network COPE, which learns the localized displacements caused by transformations within the latent space. However, jointly training COPE with the MAE leads to undesirable trivial solutions where COPE outputs collapse to an identity. To address this, we propose a loss that uses transformation-conditioned pseudo-negatives, to penalize such trivial invariant solutions. We validate PseudoNeg-MAE on shape classification and relative pose estimation tasks, where it achieves competitive performance on the ModelNet40 and ScanObjectNN datasets under challenging evaluation protocols and demonstrates superior accuracy in estimating relative poses compared to supervised methods.
3D shape completion is traditionally solved using supervised training or by distribution learning on complete shape examples. Recently self-supervised learning approaches that do not require any complete 3D shape examples have gained more interests. In this paper, we propose a non-adversarial self-supervised approach for the shape completion task. Our first finding is that completion problems can be formulated as an involutory function trivially, which implies a special constraint on the completion function G, such that G(G(X)) = X. Our second constraint on self-supervised shape completion relies on the fact that shape completion becomes easier to solve with correspondences and similarly, completion can simplify the correspondences problem. We formulate a consistency measure in the canonical space in order to supervise the completion function. We efficiently optimize the completion and correspondence modules using "freeze and alternate" strategy. The overall approach performs well for rigid shapes in a category as well as dynamic non-rigid shapes. We ablate our design choices and compare our solution against state-of-the-art methods, showing remarkable accuracy approaching supervised accuracy in some cases.
Protein function prediction is a pivotal task in drug discovery, significantly impacting the development of effective and safe therapeutics. Traditional machine learning models often struggle with the complexity and variability inherent in predicting protein functions, necessitating more sophisticated approaches. In this work, we introduce Protein-Mamba, a novel two-stage model that leverages both self-supervised learning and fine-tuning to improve protein function prediction. The pre-training stage allows the model to capture general chemical structures and relationships from large, unlabeled datasets, while the fine-tuning stage refines these insights using specific labeled datasets, resulting in superior prediction performance. Our extensive experiments demonstrate that Protein-Mamba achieves competitive performance, compared with a couple of state-of-the-art methods across a range of protein function datasets. This model's ability to effectively utilize both unlabeled and labeled data highlights the potential of self-supervised learning in advancing protein function prediction and offers a promising direction for future research in drug discovery.
Self-supervised learning (SSL) methods for large speech models have proven to be highly effective at ASR. With the interest in public deployment of large pre-trained models, there is a rising concern for unintended memorization and leakage of sensitive data points from the training data. In this paper, we apply differentially private (DP) pre-training to a SOTA Conformer-based encoder, and study its performance on a downstream ASR task assuming the fine-tuning data is public. This paper is the first to apply DP to SSL for ASR, investigating the DP noise tolerance of the BEST-RQ pre-training method. Notably, we introduce a novel variant of model pruning called gradient-based layer freezing that provides strong improvements in privacy-utility-compute trade-offs. Our approach yields a LibriSpeech test-clean/other WER (%) of 3.78/ 8.41 with ($10$, 1e^-9)-DP for extrapolation towards low dataset scales, and 2.81/ 5.89 with (10, 7.9e^-11)-DP for extrapolation towards high scales.
As per recent studies, Self-supervised learning (SSL) does not readily extend to smaller architectures. One direction to mitigate this shortcoming while simultaneously training a smaller network without labels is to adopt unsupervised knowledge distillation (UKD). Existing UKD approaches handcraft preservation worthy inter/intra sample relationships between the teacher and its student. However, this may overlook/ignore other key relationships present in the mapping of a teacher. In this paper, instead of heuristically constructing preservation worthy relationships between samples, we directly motivate the student to model the teacher's embedding manifold. If the mapped manifold is similar, all inter/intra sample relationships are indirectly conserved. We first demonstrate that prior methods cannot preserve teacher's latent manifold due to their sole reliance on $L_2$ normalised embedding features. Subsequently, we propose a simple objective to capture the lost information due to normalisation. Our proposed loss component, termed \textbf{space similarity}, motivates each dimension of a student's feature space to be similar to the corresponding dimension of its teacher. We perform extensive experiments demonstrating strong performance of our proposed approach on various benchmarks.
Multimodal clothing image editing refers to the precise adjustment and modification of clothing images using data such as textual descriptions and visual images as control conditions, which effectively improves the work efficiency of designers and reduces the threshold for user design. In this paper, we propose a new image editing method ControlEdit, which transfers clothing image editing to multimodal-guided local inpainting of clothing images. We address the difficulty of collecting real image datasets by leveraging the self-supervised learning approach. Based on this learning approach, we extend the channels of the feature extraction network to ensure consistent clothing image style before and after editing, and we design an inverse latent loss function to achieve soft control over the content of non-edited areas. In addition, we adopt Blended Latent Diffusion as the sampling method to make the editing boundaries transition naturally and enforce consistency of non-edited area content. Extensive experiments demonstrate that ControlEdit surpasses baseline algorithms in both qualitative and quantitative evaluations.
In this work, we significantly enhance masked particle modeling (MPM), a self-supervised learning scheme for constructing highly expressive representations of unordered sets relevant to developing foundation models for high-energy physics. In MPM, a model is trained to recover the missing elements of a set, a learning objective that requires no labels and can be applied directly to experimental data. We achieve significant performance improvements over previous work on MPM by addressing inefficiencies in the implementation and incorporating a more powerful decoder. We compare several pre-training tasks and introduce new reconstruction methods that utilize conditional generative models without data tokenization or discretization. We show that these new methods outperform the tokenized learning objective from the original MPM on a new test bed for foundation models for jets, which includes using a wide variety of downstream tasks relevant to jet physics, such as classification, secondary vertex finding, and track identification.
Activity cliffs, which refer to pairs of molecules that are structurally similar but show significant differences in their potency, can lead to model representation collapse and make the model challenging to distinguish them. Our research indicates that as molecular similarity increases, graph-based methods struggle to capture these nuances, whereas image-based approaches effectively retain the distinctions. Thus, we developed MaskMol, a knowledge-guided molecular image self-supervised learning framework. MaskMol accurately learns the representation of molecular images by considering multiple levels of molecular knowledge, such as atoms, bonds, and substructures. By utilizing pixel masking tasks, MaskMol extracts fine-grained information from molecular images, overcoming the limitations of existing deep learning models in identifying subtle structural changes. Experimental results demonstrate MaskMol's high accuracy and transferability in activity cliff estimation and compound potency prediction across 20 different macromolecular targets, outperforming 25 state-of-the-art deep learning and machine learning approaches. Visualization analyses reveal MaskMol's high biological interpretability in identifying activity cliff-relevant molecular substructures. Notably, through MaskMol, we identified candidate EP4 inhibitors that could be used to treat tumors. This study not only raises awareness about activity cliffs but also introduces a novel method for molecular image representation learning and virtual screening, advancing drug discovery and providing new insights into structure-activity relationships (SAR).
In this study, we investigate the application of self-supervised learning via pre-trained Long Short-Term Memory (LSTM) networks for training surface electromyography pattern recognition models (sEMG-PR) using dynamic data with transitions. While labeling such data poses challenges due to the absence of ground-truth labels during transitions between classes, self-supervised pre-training offers a way to circumvent this issue. We compare the performance of LSTMs trained with either fully-supervised or self-supervised loss to a conventional non-temporal model (LDA) on two data types: segmented ramp data (lacking transition information) and continuous dynamic data inclusive of class transitions. Statistical analysis reveals that the temporal models outperform non-temporal models when trained with continuous dynamic data. Additionally, the proposed VICReg pre-trained temporal model with continuous dynamic data significantly outperformed all other models. Interestingly, when using only ramp data, the LSTM performed worse than the LDA, suggesting potential overfitting due to the absence of sufficient dynamics. This highlights the interplay between data type and model choice. Overall, this work highlights the importance of representative dynamics in training data and the potential for leveraging self-supervised approaches to enhance sEMG-PR models.
The growing popularity of multi-channel wearable devices, such as smart glasses, has led to a surge of applications such as targeted speech recognition and enhanced hearing. However, current approaches to solve these tasks use independently trained models, which may not benefit from large amounts of unlabeled data. In this paper, we propose M-BEST-RQ, the first multi-channel speech foundation model for smart glasses, which is designed to leverage large-scale self-supervised learning (SSL) in an array-geometry agnostic approach. While prior work on multi-channel speech SSL only evaluated on simulated settings, we curate a suite of real downstream tasks to evaluate our model, namely (i) conversational automatic speech recognition (ASR), (ii) spherical active source localization, and (iii) glasses wearer voice activity detection, which are sourced from the MMCSG and EasyCom datasets. We show that a general-purpose M-BEST-RQ encoder is able to match or surpass supervised models across all tasks. For the conversational ASR task in particular, using only 8 hours of labeled speech, our model outperforms a supervised ASR baseline that is trained on 2000 hours of labeled data, which demonstrates the effectiveness of our approach.
This study explores using embedding rank as an unsupervised evaluation metric for general-purpose speech encoders trained via self-supervised learning (SSL). Traditionally, assessing the performance of these encoders is resource-intensive and requires labeled data from the downstream tasks. Inspired by the vision domain, where embedding rank has shown promise for evaluating image encoders without tuning on labeled downstream data, this work examines its applicability in the speech domain, considering the temporal nature of the signals. The findings indicate rank correlates with downstream performance within encoder layers across various downstream tasks and for in- and out-of-domain scenarios. However, rank does not reliably predict the best-performing layer for specific downstream tasks, as lower-ranked layers can outperform higher-ranked ones. Despite this limitation, the results suggest that embedding rank can be a valuable tool for monitoring training progress in SSL speech models, offering a less resource-demanding alternative to traditional evaluation methods.
Side-scan sonar (SSS) imagery presents unique challenges in the classification of man-made objects on the seafloor due to the complex and varied underwater environments. Historically, experts have manually interpreted SSS images, relying on conventional machine learning techniques with hand-crafted features. While Convolutional Neural Networks (CNNs) significantly advanced automated classification in this domain, they often fall short when dealing with diverse seafloor textures, such as rocky or ripple sand bottoms, where false positive rates may increase. Recently, Vision Transformers (ViTs) have shown potential in addressing these limitations by utilizing a self-attention mechanism to capture global information in image patches, offering more flexibility in processing spatial hierarchies. This paper rigorously compares the performance of ViT models alongside commonly used CNN architectures, such as ResNet and ConvNext, for binary classification tasks in SSS imagery. The dataset encompasses diverse geographical seafloor types and is balanced between the presence and absence of man-made objects. ViT-based models exhibit superior classification performance across f1-score, precision, recall, and accuracy metrics, although at the cost of greater computational resources. CNNs, with their inductive biases, demonstrate better computational efficiency, making them suitable for deployment in resource-constrained environments like underwater vehicles. Future research directions include exploring self-supervised learning for ViTs and multi-modal fusion to further enhance performance in challenging underwater environments.
Imitation learning has proven to be a powerful tool for training complex visuomotor policies. However, current methods often require hundreds to thousands of expert demonstrations to handle high-dimensional visual observations. A key reason for this poor data efficiency is that visual representations are predominantly either pretrained on out-of-domain data or trained directly through a behavior cloning objective. In this work, we present DynaMo, a new in-domain, self-supervised method for learning visual representations. Given a set of expert demonstrations, we jointly learn a latent inverse dynamics model and a forward dynamics model over a sequence of image embeddings, predicting the next frame in latent space, without augmentations, contrastive sampling, or access to ground truth actions. Importantly, DynaMo does not require any out-of-domain data such as Internet datasets or cross-embodied datasets. On a suite of six simulated and real environments, we show that representations learned with DynaMo significantly improve downstream imitation learning performance over prior self-supervised learning objectives, and pretrained representations. Gains from using DynaMo hold across policy classes such as Behavior Transformer, Diffusion Policy, MLP, and nearest neighbors. Finally, we ablate over key components of DynaMo and measure its impact on downstream policy performance. Robot videos are best viewed at https://dynamo-ssl.github.io