Recently, there have been efforts to improve the performance in sign language recognition by designing self-supervised learning methods. However, these methods capture limited information from sign pose data in a frame-wise learning manner, leading to sub-optimal solutions. To this end, we propose a simple yet effective self-supervised contrastive learning framework to excavate rich context via spatial-temporal consistency from two distinct perspectives and learn instance discriminative representation for sign language recognition. On one hand, since the semantics of sign language are expressed by the cooperation of fine-grained hands and coarse-grained trunks, we utilize both granularity information and encode them into latent spaces. The consistency between hand and trunk features is constrained to encourage learning consistent representation of instance samples. On the other hand, inspired by the complementary property of motion and joint modalities, we first introduce first-order motion information into sign language modeling. Additionally, we further bridge the interaction between the embedding spaces of both modalities, facilitating bidirectional knowledge transfer to enhance sign language representation. Our method is evaluated with extensive experiments on four public benchmarks, and achieves new state-of-the-art performance with a notable margin. The source code is publicly available at https://github.com/sakura/Code.
Deep Learning is often depicted as a trio of data-architecture-loss. Yet, recent Self Supervised Learning (SSL) solutions have introduced numerous additional design choices, e.g., a projector network, positive views, or teacher-student networks. These additions pose two challenges. First, they limit the impact of theoretical studies that often fail to incorporate all those intertwined designs. Second, they slow-down the deployment of SSL methods to new domains as numerous hyper-parameters need to be carefully tuned. In this study, we bring forward the surprising observation that--at least for pretraining datasets of up to a few hundred thousands samples--the additional designs introduced by SSL do not contribute to the quality of the learned representations. That finding not only provides legitimacy to existing theoretical studies, but also simplifies the practitioner's path to SSL deployment in numerous small and medium scale settings. Our finding answers a long-lasting question: the often-experienced sensitivity to training settings and hyper-parameters encountered in SSL come from their design, rather than the absence of supervised guidance.
Discrete audio tokens have recently gained attention for their potential to bridge the gap between audio and language processing. Ideal audio tokens must preserve content, paralinguistic elements, speaker identity, and many other audio details. Current audio tokenization methods fall into two categories: Semantic tokens, acquired through quantization of Self-Supervised Learning (SSL) models, and Neural compression-based tokens (codecs). Although previous studies have benchmarked codec models to identify optimal configurations, the ideal setup for quantizing pretrained SSL models remains unclear. This paper explores the optimal configuration of semantic tokens across discriminative and generative tasks. We propose a scalable solution to train a universal vocoder across multiple SSL layers. Furthermore, an attention mechanism is employed to identify task-specific influential layers, enhancing the adaptability and performance of semantic tokens in diverse audio applications.
Foundation models (FMs) are large-scale deep learning models trained on massive datasets, often using self-supervised learning techniques. These models serve as a versatile base for a wide range of downstream tasks, including those in medicine and healthcare. FMs have demonstrated remarkable success across multiple healthcare domains. However, existing surveys in this field do not comprehensively cover all areas where FMs have made significant strides. In this survey, we present a comprehensive review of FMs in medicine, focusing on their evolution, learning strategies, flagship models, applications, and associated challenges. We examine how prominent FMs, such as the BERT and GPT families, are transforming various aspects of healthcare, including clinical large language models, medical image analysis, and omics research. Additionally, we provide a detailed taxonomy of FM-enabled healthcare applications, spanning clinical natural language processing, medical computer vision, graph learning, and other biology- and omics- related tasks. Despite the transformative potentials of FMs, they also pose unique challenges. This survey delves into these challenges and highlights open research questions and lessons learned to guide researchers and practitioners. Our goal is to provide valuable insights into the capabilities of FMs in health, facilitating responsible deployment and mitigating associated risks.
Automatic speech recognition (ASR) systems often falter while processing stuttering-related disfluencies -- such as involuntary blocks and word repetitions -- yielding inaccurate transcripts. A critical barrier to progress is the scarcity of large, annotated disfluent speech datasets. Therefore, we present an inclusive ASR design approach, leveraging large-scale self-supervised learning on standard speech followed by targeted fine-tuning and data augmentation on a smaller, curated dataset of disfluent speech. Our data augmentation technique enriches training datasets with various disfluencies, enhancing ASR processing of these speech patterns. Results show that fine-tuning wav2vec 2.0 with even a relatively small, labeled dataset, alongside data augmentation, can significantly reduce word error rates for disfluent speech. Our approach not only advances ASR inclusivity for people who stutter, but also paves the way for ASRs that can accommodate wider speech variations.
Real-time bioaerosol monitoring is improving the quality of life for people affected by allergies, but it often relies on deep-learning models which pose challenges for widespread adoption. These models are typically trained in a supervised fashion and require considerable effort to produce large amounts of annotated data, an effort that must be repeated for new particles, geographical regions, or measurement systems. In this work, we show that self-supervised learning and few-shot learning can be combined to classify holographic images of bioaerosol particles using a large collection of unlabelled data and only a few examples for each particle type. We first demonstrate that self-supervision on pictures of unidentified particles from ambient air measurements enhances identification even when labelled data is abundant. Most importantly, it greatly improves few-shot classification when only a handful of labelled images are available. Our findings suggest that real-time bioaerosol monitoring workflows can be substantially optimized, and the effort required to adapt models for different situations considerably reduced.
We consider the problem of \textit{true} open-world semi-supervised node classification, in which nodes in a graph either belong to known or new classes, with the latter not present during training. Existing methods detect and reject new classes but fail to distinguish between different new classes. We adapt existing methods and show they do not solve the problem sufficiently. We introduce a novel end-to-end approach for classification into known classes and new classes based on class prototypes, which we call Prototypical Open-World Learning for Node Classification (POWN). Our method combines graph semi-supervised learning, self-supervised learning, and pseudo-labeling to learn prototype representations of new classes in a zero-shot way. In contrast to existing solutions from the vision domain, POWN does not require data augmentation techniques for node classification. Experiments on benchmark datasets demonstrate the effectiveness of POWN, where it outperforms baselines by up to $20\%$ accuracy on the small and up to $30\%$ on the large datasets. Source code is available at https://github.com/Bobowner/POWN.
State-of-the-art 3D object detectors are often trained on massive labeled datasets. However, annotating 3D bounding boxes remains prohibitively expensive and time-consuming, particularly for LiDAR. Instead, recent works demonstrate that self-supervised pre-training with unlabeled data can improve detection accuracy with limited labels. Contemporary methods adapt best-practices for self-supervised learning from the image domain to point clouds (such as contrastive learning). However, publicly available 3D datasets are considerably smaller and less diverse than those used for image-based self-supervised learning, limiting their effectiveness. We do note, however, that such 3D data is naturally collected in a multimodal fashion, often paired with images. Rather than pre-training with only self-supervised objectives, we argue that it is better to bootstrap point cloud representations using image-based foundation models trained on internet-scale data. Specifically, we propose a shelf-supervised approach (e.g. supervised with off-the-shelf image foundation models) for generating zero-shot 3D bounding boxes from paired RGB and LiDAR data. Pre-training 3D detectors with such pseudo-labels yields significantly better semi-supervised detection accuracy than prior self-supervised pretext tasks. Importantly, we show that image-based shelf-supervision is helpful for training LiDAR-only, RGB-only and multi-modal (RGB + LiDAR) detectors. We demonstrate the effectiveness of our approach on nuScenes and WOD, significantly improving over prior work in limited data settings. Our code is available at https://github.com/meharkhurana03/cm3d
Polyps are early cancer indicators, so assessing occurrences of polyps and their removal is critical. They are observed through a colonoscopy screening procedure that generates a stream of video frames. Segmenting polyps in their natural video screening procedure has several challenges, such as the co-existence of imaging artefacts, motion blur, and floating debris. Most existing polyp segmentation algorithms are developed on curated still image datasets that do not represent real-world colonoscopy. Their performance often degrades on video data. We propose a video polyp segmentation method that performs self-supervised learning as an auxiliary task and a spatial-temporal self-attention mechanism for improved representation learning. Our end-to-end configuration and joint optimisation of losses enable the network to learn more discriminative contextual features in videos. Our experimental results demonstrate an improvement with respect to several state-of-the-art (SOTA) methods. Our ablation study also confirms that the choice of the proposed joint end-to-end training improves network accuracy by over 3% and nearly 10% on both the Dice similarity coefficient and intersection-over-union compared to the recently proposed method PNS+ and Polyp-PVT, respectively. Results on previously unseen video data indicate that the proposed method generalises.
Speaker Change Detection (SCD) is to identify boundaries among speakers in a conversation. Motivated by the success of fine-tuning wav2vec 2.0 models for the SCD task, a further investigation of self-supervised learning (SSL) features for SCD is conducted in this work. Specifically, an SCD model, named SCDNet, is proposed. With this model, various state-of-the-art SSL models, including Hubert, wav2vec 2.0, and WavLm are investigated. To discern the most potent layer of SSL models for SCD, a learnable weighting method is employed to analyze the effectiveness of intermediate representations. Additionally, a fine-tuning-based approach is also implemented to further compare the characteristics of SSL models in the SCD task. Furthermore, a contrastive learning method is proposed to mitigate the overfitting tendencies in the training of both the fine-tuning-based method and SCDNet. Experiments showcase the superiority of WavLm in the SCD task and also demonstrate the good design of SCDNet.
Self-supervised learning (SSL)-based speech models are extensively used for full-stack speech processing. However, it has been observed that improving SSL-based speech representations using unlabeled speech for content-related tasks is challenging and computationally expensive. Recent attempts have been made to address this issue with cost-effective self-supervised fine-tuning (SSFT) approaches. Continuing in this direction, a cost-effective SSFT method named "LASER: Learning by Aligning Self-supervised Representations" is presented. LASER is based on the soft-DTW alignment loss with temporal regularisation term. Experiments are conducted with HuBERT and WavLM models and evaluated on the SUPERB benchmark for two content-related tasks: automatic speech recognition (ASR) and phoneme recognition (PR). A relative improvement of 3.7% and 8.2% for HuBERT, and 4.1% and 11.7% for WavLM are observed, for the ASR and PR tasks respectively, with only < 3 hours of fine-tuning on a single GPU.
Self-supervised learning (SSL) representations from massively multilingual models offer a promising solution for low-resource language speech tasks. Despite advancements, language adaptation in TTS systems remains an open problem. This paper explores the language adaptation capability of ZMM-TTS, a recent SSL-based multilingual TTS system proposed in our previous work. We conducted experiments on 12 languages using limited data with various fine-tuning configurations. We demonstrate that the similarity in phonetics between the pre-training and target languages, as well as the language category, affects the target language's adaptation performance. Additionally, we find that the fine-tuning dataset size and number of speakers influence adaptability. Surprisingly, we also observed that using paired data for fine-tuning is not always optimal compared to audio-only data. Beyond speech intelligibility, our analysis covers speaker similarity, language identification, and predicted MOS.
ML-SUPERB evaluates self-supervised learning (SSL) models on the tasks of language identification and automatic speech recognition (ASR). This benchmark treats the models as feature extractors and uses a single shallow downstream model, which can be fine-tuned for a downstream task. However, real-world use cases may require different configurations. This paper presents ML-SUPERB~2.0, which is a new benchmark for evaluating pre-trained SSL and supervised speech models across downstream models, fine-tuning setups, and efficient model adaptation approaches. We find performance improvements over the setup of ML-SUPERB. However, performance depends on the downstream model design. Also, we find large performance differences between languages and datasets, suggesting the need for more targeted approaches to improve multilingual ASR performance.
Self-Supervised learning (SSL) with Joint-Embedding Architectures (JEA) has led to outstanding performances. All instantiations of this paradigm were trained using strong and well-established hand-crafted data augmentations, leading to the general belief that they are required for the proper training and performance of such models. On the other hand, generative reconstruction-based models such as BEIT and MAE or Joint-Embedding Predictive Architectures such as I-JEPA have shown strong performance without using data augmentations except masking. In this work, we challenge the importance of invariance and data-augmentation in JEAs at scale. By running a case-study on a recent SSL foundation model - DINOv2 - we show that strong image representations can be obtained with JEAs and only cropping without resizing provided the training data is large enough, reaching state-of-the-art results and using the least amount of augmentation in the literature. Through this study, we also discuss the impact of compute constraints on the outcomes of experimental deep learning research, showing that they can lead to very different conclusions.
This work does not introduce a new method. Instead, we present an interesting finding that questions the necessity of the inductive bias -- locality in modern computer vision architectures. Concretely, we find that vanilla Transformers can operate by directly treating each individual pixel as a token and achieve highly performant results. This is substantially different from the popular design in Vision Transformer, which maintains the inductive bias from ConvNets towards local neighborhoods (e.g. by treating each 16x16 patch as a token). We mainly showcase the effectiveness of pixels-as-tokens across three well-studied tasks in computer vision: supervised learning for object classification, self-supervised learning via masked autoencoding, and image generation with diffusion models. Although directly operating on individual pixels is less computationally practical, we believe the community must be aware of this surprising piece of knowledge when devising the next generation of neural architectures for computer vision.
Maximum Manifold Capacity Representations (MMCR) is a recent multi-view self-supervised learning (MVSSL) method that matches or surpasses other leading MVSSL methods. MMCR is intriguing because it does not fit neatly into any of the commonplace MVSSL lineages, instead originating from a statistical mechanical perspective on the linear separability of data manifolds. In this paper, we seek to improve our understanding and our utilization of MMCR. To better understand MMCR, we leverage tools from high dimensional probability to demonstrate that MMCR incentivizes alignment and uniformity of learned embeddings. We then leverage tools from information theory to show that such embeddings maximize a well-known lower bound on mutual information between views, thereby connecting the geometric perspective of MMCR to the information-theoretic perspective commonly discussed in MVSSL. To better utilize MMCR, we mathematically predict and experimentally confirm non-monotonic changes in the pretraining loss akin to double descent but with respect to atypical hyperparameters. We also discover compute scaling laws that enable predicting the pretraining loss as a function of gradients steps, batch size, embedding dimension and number of views. We then show that MMCR, originally applied to image data, is performant on multimodal image-text data. By more deeply understanding the theoretical and empirical behavior of MMCR, our work reveals insights on improving MVSSL methods.
Multimedia recommendation, which incorporates various modalities (e.g., images, texts, etc.) into user or item representation to improve recommendation quality, and self-supervised learning carries multimedia recommendation to a plateau of performance, because of its superior performance in aligning different modalities. However, more and more research finds that aligning all modal representations is suboptimal because it damages the unique attributes of each modal. These studies use subtraction and orthogonal constraints in geometric space to learn unique parts. However, our rigorous analysis reveals the flaws in this approach, such as that subtraction does not necessarily yield the desired modal-unique and that orthogonal constraints are ineffective in user and item high-dimensional representation spaces. To make up for the previous weaknesses, we propose Separate Learning (SEA) for multimedia recommendation, which mainly includes mutual information view of modal-unique and -generic learning. Specifically, we first use GNN to learn the representations of users and items in different modalities and split each modal representation into generic and unique parts. We employ contrastive log-ratio upper bound to minimize the mutual information between the general and unique parts within the same modality, to distance their representations, thus learning modal-unique features. Then, we design Solosimloss to maximize the lower bound of mutual information, to align the general parts of different modalities, thus learning more high-quality modal-generic features. Finally, extensive experiments on three datasets demonstrate the effectiveness and generalization of our proposed framework. The code is available at SEA and the full training record of the main experiment.
Sustainable artificial intelligence focuses on data, hardware, and algorithms to make machine learning models more environmentally responsible. In particular, machine learning models for speech representations are computationally expensive, generating environmental concerns because of their high energy consumption. Thus, we propose a sustainable self-supervised model to learn speech representation, combining optimizations in neural layers and training to reduce computing costs. The proposed model improves over a resource-efficient baseline, reducing both memory usage and computing cost estimations. It pretrains using a single GPU in less than a day. On top of that, it improves the error rate performance of the baseline in downstream task evaluations. When comparing it to large speech representation approaches, there is an order of magnitude reduction in memory usage, while computing cost reductions represent almost three orders of magnitude improvement.
Foundation Models (FMs) serve as a general class for the development of artificial intelligence systems, offering broad potential for generalization across a spectrum of downstream tasks. Despite extensive research into self-supervised learning as the cornerstone of FMs, several outstanding issues persist in Graph Foundation Models that rely on graph self-supervised learning, namely: 1) Homogenization. The extent of generalization capability on downstream tasks remains unclear. 2) Scalability. It is unknown how effectively these models can scale to large datasets. 3) Efficiency. The training time and memory usage of these models require evaluation. 4) Training Stop Criteria. Determining the optimal stopping strategy for pre-training across multiple tasks to maximize performance on downstream tasks. To address these questions, we have constructed a rigorous benchmark that thoroughly analyzes and studies the generalization and scalability of self-supervised Graph Neural Network (GNN) models. Regarding generalization, we have implemented and compared the performance of various self-supervised GNN models, trained to generate node representations, across tasks such as node classification, link prediction, and node clustering. For scalability, we have compared the performance of various models after training using full-batch and mini-batch strategies. Additionally, we have assessed the training efficiency of these models by conducting experiments to test their GPU memory usage and throughput. Through these experiments, we aim to provide insights to motivate future research. The code for this benchmark is publicly available at https://github.com/NYUSHCS/GraphFM.
Deoccluding the hidden portions of objects in a scene is a formidable task, particularly when addressing real-world scenes. In this paper, we present a new self-supervised PArallel visible-to-COmplete diffusion framework, named PACO, a foundation model for object-level scene deocclusion. Leveraging the rich prior of pre-trained models, we first design the parallel variational autoencoder, which produces a full-view feature map that simultaneously encodes multiple complete objects, and the visible-to-complete latent generator, which learns to implicitly predict the full-view feature map from partial-view feature map and text prompts extracted from the incomplete objects in the input image. To train PACO, we create a large-scale dataset with 500k samples to enable self-supervised learning, avoiding tedious annotations of the amodal masks and occluded regions. At inference, we devise a layer-wise deocclusion strategy to improve efficiency while maintaining the deocclusion quality. Extensive experiments on COCOA and various real-world scenes demonstrate the superior capability of PACO for scene deocclusion, surpassing the state of the arts by a large margin. Our method can also be extended to cross-domain scenes and novel categories that are not covered by the training set. Further, we demonstrate the deocclusion applicability of PACO in single-view 3D scene reconstruction and object recomposition.