Uniformity plays an important role in evaluating learned representations, providing insights into self-supervised learning. In our quest for effective uniformity metrics, we pinpoint four principled properties that such metrics should possess. Namely, an effective uniformity metric should remain invariant to instance permutations and sample replications while accurately capturing feature redundancy and dimensional collapse. Surprisingly, we find that the uniformity metric proposed by \citet{Wang2020UnderstandingCR} fails to satisfy the majority of these properties. Specifically, their metric is sensitive to sample replications, and can not account for feature redundancy and dimensional collapse correctly. To overcome these limitations, we introduce a new uniformity metric based on the Wasserstein distance, which satisfies all the aforementioned properties. Integrating this new metric in existing self-supervised learning methods effectively mitigates dimensional collapse and consistently improves their performance on downstream tasks involving CIFAR-10 and CIFAR-100 datasets. Code is available at \url{https://github.com/statsle/WassersteinSSL}.
With the increasing multimedia information, multimodal recommendation has received extensive attention. It utilizes multimodal information to alleviate the data sparsity problem in recommendation systems, thus improving recommendation accuracy. However, the reliance on labeled data severely limits the performance of multimodal recommendation models. Recently, self-supervised learning has been used in multimodal recommendations to mitigate the label sparsity problem. Nevertheless, the state-of-the-art methods cannot avoid the modality noise when aligning multimodal information due to the large differences in the distributions of different modalities. To this end, we propose a Multi-level sElf-supervised learNing for mulTimOdal Recommendation (MENTOR) method to address the label sparsity problem and the modality alignment problem. Specifically, MENTOR first enhances the specific features of each modality using the graph convolutional network (GCN) and fuses the visual and textual modalities. It then enhances the item representation via the item semantic graph for all modalities, including the fused modality. Then, it introduces two multilevel self-supervised tasks: the multilevel cross-modal alignment task and the general feature enhancement task. The multilevel cross-modal alignment task aligns each modality under the guidance of the ID embedding from multiple levels while maintaining the historical interaction information. The general feature enhancement task enhances the general feature from both the graph and feature perspectives to improve the robustness of our model. Extensive experiments on three publicly available datasets demonstrate the effectiveness of our method. Our code is publicly available at https://github.com/Jinfeng-Xu/MENTOR.
Recently, the advancement of self-supervised learning techniques, like masked autoencoders (MAE), has greatly influenced visual representation learning for images and videos. Nevertheless, it is worth noting that the predominant approaches in existing masked image / video modeling rely excessively on resource-intensive vision transformers (ViTs) as the feature encoder. In this paper, we propose a new approach termed as \textbf{VideoMAC}, which combines video masked autoencoders with resource-friendly ConvNets. Specifically, VideoMAC employs symmetric masking on randomly sampled pairs of video frames. To prevent the issue of mask pattern dissipation, we utilize ConvNets which are implemented with sparse convolutional operators as encoders. Simultaneously, we present a simple yet effective masked video modeling (MVM) approach, a dual encoder architecture comprising an online encoder and an exponential moving average target encoder, aimed to facilitate inter-frame reconstruction consistency in videos. Additionally, we demonstrate that VideoMAC, empowering classical (ResNet) / modern (ConvNeXt) convolutional encoders to harness the benefits of MVM, outperforms ViT-based approaches on downstream tasks, including video object segmentation (+\textbf{5.2\%} / \textbf{6.4\%} $\mathcal{J}\&\mathcal{F}$), body part propagation (+\textbf{6.3\%} / \textbf{3.1\%} mIoU), and human pose tracking (+\textbf{10.2\%} / \textbf{11.1\%} [email omitted]).
Sentence Representation Learning (SRL) is a crucial task in Natural Language Processing (NLP), where contrastive Self-Supervised Learning (SSL) is currently a mainstream approach. However, the reasons behind its remarkable effectiveness remain unclear. Specifically, many studies have investigated the similarities between contrastive and non-contrastive SSL from a theoretical perspective. Such similarities can be verified in classification tasks, where the two approaches achieve comparable performance. But in ranking tasks (i.e., Semantic Textual Similarity (STS) in SRL), contrastive SSL significantly outperforms non-contrastive SSL. Therefore, two questions arise: First, *what commonalities enable various contrastive losses to achieve superior performance in STS?* Second, *how can we make non-contrastive SSL also effective in STS?* To address these questions, we start from the perspective of gradients and discover that four effective contrastive losses can be integrated into a unified paradigm, which depends on three components: the **Gradient Dissipation**, the **Weight**, and the **Ratio**. Then, we conduct an in-depth analysis of the roles these components play in optimization and experimentally demonstrate their significance for model performance. Finally, by adjusting these components, we enable non-contrastive SSL to achieve outstanding performance in STS.
Self-Supervised Learning (SSL) has demonstrated promising results in 3D medical image analysis. However, the lack of high-level semantics in pre-training still heavily hinders the performance of downstream tasks. We observe that 3D medical images contain relatively consistent contextual position information, i.e., consistent geometric relations between different organs, which leads to a potential way for us to learn consistent semantic representations in pre-training. In this paper, we propose a simple-yet-effective Volume Contrast (VoCo) framework to leverage the contextual position priors for pre-training. Specifically, we first generate a group of base crops from different regions while enforcing feature discrepancy among them, where we employ them as class assignments of different regions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which region) by contrasting their similarity to different base crops, which can be seen as predicting contextual positions of different sub-volumes. Through this pretext task, VoCo implicitly encodes the contextual position priors into model representations without the guidance of annotations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experimental results on six downstream tasks demonstrate the superior effectiveness of VoCo. Code will be available at https://github.com/Luffy03/VoCo.
Multi-pitch estimation is a decades-long research problem involving the detection of pitch activity associated with concurrent musical events within multi-instrument mixtures. Supervised learning techniques have demonstrated solid performance on more narrow characterizations of the task, but suffer from limitations concerning the shortage of large-scale and diverse polyphonic music datasets with multi-pitch annotations. We present a suite of self-supervised learning objectives for multi-pitch estimation, which encourage the concentration of support around harmonics, invariance to timbral transformations, and equivariance to geometric transformations. These objectives are sufficient to train an entirely convolutional autoencoder to produce multi-pitch salience-grams directly, without any fine-tuning. Despite training exclusively on a collection of synthetic single-note audio samples, our fully self-supervised framework generalizes to polyphonic music mixtures, and achieves performance comparable to supervised models trained on conventional multi-pitch datasets.
This paper focuses on semi-supervised crowd counting, where only a small portion of the training data are labeled. We formulate the pixel-wise density value to regress as a probability distribution, instead of a single deterministic value. On this basis, we propose a semi-supervised crowd-counting model. Firstly, we design a pixel-wise distribution matching loss to measure the differences in the pixel-wise density distributions between the prediction and the ground truth; Secondly, we enhance the transformer decoder by using density tokens to specialize the forwards of decoders w.r.t. different density intervals; Thirdly, we design the interleaving consistency self-supervised learning mechanism to learn from unlabeled data efficiently. Extensive experiments on four datasets are performed to show that our method clearly outperforms the competitors by a large margin under various labeled ratio settings. Code will be released at https://github.com/LoraLinH/Semi-supervised-Counting-via-Pixel-by-pixel-Density-Distribution-Modelling.
Last couple of years have witnessed a tremendous progress in self-supervised learning (SSL), the success of which can be attributed to the introduction of useful inductive biases in the learning process to learn meaningful visual representations while avoiding collapse. These inductive biases and constraints manifest themselves in the form of different optimization formulations in the SSL techniques, e.g. by utilizing negative examples in a contrastive formulation, or exponential moving average and predictor in BYOL and SimSiam. In this paper, we provide a framework to explain the stability mechanism of these different SSL techniques: i) we discuss the working mechanism of contrastive techniques like SimCLR, non-contrastive techniques like BYOL, SWAV, SimSiam, Barlow Twins, and DINO; ii) we provide an argument that despite different formulations these methods implicitly optimize a similar objective function, i.e. minimizing the magnitude of the expected representation over all data samples, or the mean of the data distribution, while maximizing the magnitude of the expected representation of individual samples over different data augmentations; iii) we provide mathematical and empirical evidence to support our framework. We formulate different hypotheses and test them using the Imagenet100 dataset.
Self-supervised learning methods based on data augmentations, such as SimCLR, BYOL, or DINO, allow obtaining semantically meaningful representations of image datasets and are widely used prior to supervised fine-tuning. A recent self-supervised learning method, $t$-SimCNE, uses contrastive learning to directly train a 2D representation suitable for visualisation. When applied to natural image datasets, $t$-SimCNE yields 2D visualisations with semantically meaningful clusters. In this work, we used $t$-SimCNE to visualise medical image datasets, including examples from dermatology, histology, and blood microscopy. We found that increasing the set of data augmentations to include arbitrary rotations improved the results in terms of class separability, compared to data augmentations used for natural images. Our 2D representations show medically relevant structures and can be used to aid data exploration and annotation, improving on common approaches for data visualisation.
Conventional deep learning models deal with images one-by-one, requiring costly and time-consuming expert labeling in the field of medical imaging, and domain-specific restriction limits model generalizability. Visual in-context learning (ICL) is a new and exciting area of research in computer vision. Unlike conventional deep learning, ICL emphasizes the model's ability to adapt to new tasks based on given examples quickly. Inspired by MAE-VQGAN, we proposed a new simple visual ICL method called SimICL, combining visual ICL pairing images with masked image modeling (MIM) designed for self-supervised learning. We validated our method on bony structures segmentation in a wrist ultrasound (US) dataset with limited annotations, where the clinical objective was to segment bony structures to help with further fracture detection. We used a test set containing 3822 images from 18 patients for bony region segmentation. SimICL achieved an remarkably high Dice coeffient (DC) of 0.96 and Jaccard Index (IoU) of 0.92, surpassing state-of-the-art segmentation and visual ICL models (a maximum DC 0.86 and IoU 0.76), with SimICL DC and IoU increasing up to 0.10 and 0.16. This remarkably high agreement with limited manual annotations indicates SimICL could be used for training AI models even on small US datasets. This could dramatically decrease the human expert time required for image labeling compared to conventional approaches, and enhance the real-world use of AI assistance in US image analysis.
Self-supervised learning excels in learning representations from large amounts of unlabeled data, demonstrating success across multiple data modalities. Yet, extending self-supervised learning to new modalities is non-trivial because the specifics of existing methods are tailored to each domain, such as domain-specific augmentations which reflect the invariances in the target task. While masked modeling is promising as a domain-agnostic framework for self-supervised learning because it does not rely on input augmentations, its mask sampling procedure remains domain-specific. We present Self-guided Masked Autoencoders (SMA), a fully domain-agnostic masked modeling method. SMA trains an attention based model using a masked modeling objective, by learning masks to sample without any domain-specific assumptions. We evaluate SMA on three self-supervised learning benchmarks in protein biology, chemical property prediction, and particle physics. We find SMA is capable of learning representations without domain-specific knowledge and achieves state-of-the-art performance on these three benchmarks.
In recent years, self-supervised learning has emerged as a powerful tool to harness abundant unlabelled data for representation learning and has been broadly adopted in diverse areas. However, when applied to molecular representation learning (MRL), prevailing techniques such as masked sub-unit reconstruction often fall short, due to the high degree of freedom in the possible combinations of atoms within molecules, which brings insurmountable complexity to the masking-reconstruction paradigm. To tackle this challenge, we introduce REMO, a self-supervised learning framework that takes advantage of well-defined atom-combination rules in common chemistry. Specifically, REMO pre-trains graph/Transformer encoders on 1.7 million known chemical reactions in the literature. We propose two pre-training objectives: Masked Reaction Centre Reconstruction (MRCR) and Reaction Centre Identification (RCI). REMO offers a novel solution to MRL by exploiting the underlying shared patterns in chemical reactions as \textit{context} for pre-training, which effectively infers meaningful representations of common chemistry knowledge. Such contextual representations can then be utilized to support diverse downstream molecular tasks with minimum finetuning, such as affinity prediction and drug-drug interaction prediction. Extensive experimental results on MoleculeACE, ACNet, drug-drug interaction (DDI), and reaction type classification show that across all tested downstream tasks, REMO outperforms the standard baseline of single-molecule masked modeling used in current MRL. Remarkably, REMO is the pioneering deep learning model surpassing fingerprint-based methods in activity cliff benchmarks.
Foundation models in speech are often trained using many GPUs, which implicitly leads to large effective batch sizes. In this paper we study the effect of batch size on pre-training, both in terms of statistics that can be monitored during training, and in the effect on the performance of a downstream fine-tuning task. By using batch sizes varying from 87.5 seconds to 80 minutes of speech we show that, for a fixed amount of iterations, larger batch sizes result in better pre-trained models. However, there is lower limit for stability, and an upper limit for effectiveness. We then show that the quality of the pre-trained model depends mainly on the amount of speech data seen during training, i.e., on the product of batch size and number of iterations. All results are produced with an independent implementation of the wav2vec 2.0 architecture, which to a large extent reproduces the results of the original work (arXiv:2006.11477). Our extensions can help researchers choose effective operating conditions when studying self-supervised learning in speech, and hints towards benchmarking self-supervision with a fixed amount of seen data. Code and model checkpoints are available at https://github.com/nikvaessen/w2v2-batch-size.
Foundation models like ChatGPT and GPT-4 have revolutionized artificial intelligence, exhibiting remarkable abilities to generalize across a wide array of tasks and applications beyond their initial training objectives. However, graph learning has predominantly focused on single-graph models, tailored to specific tasks or datasets, lacking the ability to transfer learned knowledge to different domains. This limitation stems from the inherent complexity and diversity of graph structures, along with the different feature and label spaces specific to graph data. In this paper, we recognize text as an effective unifying medium and employ Text-Attributed Graphs (TAGs) to leverage this potential. We present our UniGraph framework, designed to learn a foundation model for TAGs, which is capable of generalizing to unseen graphs and tasks across diverse domains. Unlike single-graph models that use pre-computed node features of varying dimensions as input, our approach leverages textual features for unifying node representations, even for graphs such as molecular graphs that do not naturally have textual features. We propose a novel cascaded architecture of Language Models (LMs) and Graph Neural Networks (GNNs) as backbone networks. Additionally, we propose the first pre-training algorithm specifically designed for large-scale self-supervised learning on TAGs, based on Masked Graph Modeling. We introduce graph instruction tuning using Large Language Models (LLMs) to enable zero-shot prediction ability. Our comprehensive experiments across various graph learning tasks and domains demonstrate the model's effectiveness in self-supervised representation learning on unseen graphs, few-shot in-context transfer, and zero-shot transfer, even surpassing or matching the performance of GNNs that have undergone supervised training on target datasets.
Speech emotion recognition (SER) is a pivotal technology for human-computer interaction systems. However, 80.77% of SER papers yield results that cannot be reproduced. We develop EMO-SUPERB, short for EMOtion Speech Universal PERformance Benchmark, which aims to enhance open-source initiatives for SER. EMO-SUPERB includes a user-friendly codebase to leverage 15 state-of-the-art speech self-supervised learning models (SSLMs) for exhaustive evaluation across six open-source SER datasets. EMO-SUPERB streamlines result sharing via an online leaderboard, fostering collaboration within a community-driven benchmark and thereby enhancing the development of SER. On average, 2.58% of annotations are annotated using natural language. SER relies on classification models and is unable to process natural languages, leading to the discarding of these valuable annotations. We prompt ChatGPT to mimic annotators, comprehend natural language annotations, and subsequently re-label the data. By utilizing labels generated by ChatGPT, we consistently achieve an average relative gain of 3.08% across all settings.
To detect anomalies in real-world graphs, such as social, email, and financial networks, various approaches have been developed. While they typically assume static input graphs, most real-world graphs grow over time, naturally represented as edge streams. In this context, we aim to achieve three goals: (a) instantly detecting anomalies as they occur, (b) adapting to dynamically changing states, and (c) handling the scarcity of dynamic anomaly labels. In this paper, we propose SLADE (Self-supervised Learning for Anomaly Detection in Edge Streams) for rapid detection of dynamic anomalies in edge streams, without relying on labels. SLADE detects the shifts of nodes into abnormal states by observing deviations in their interaction patterns over time. To this end, it trains a deep neural network to perform two self-supervised tasks: (a) minimizing drift in node representations and (b) generating long-term interaction patterns from short-term ones. Failure in these tasks for a node signals its deviation from the norm. Notably, the neural network and tasks are carefully designed so that all required operations can be performed in constant time (w.r.t. the graph size) in response to each new edge in the input stream. In dynamic anomaly detection across four real-world datasets, SLADE outperforms nine competing methods, even those leveraging label supervision.
Electrocardiograms (ECG) are widely employed as a diagnostic tool for monitoring electrical signals originating from a heart. Recent machine learning research efforts have focused on the application of screening various diseases using ECG signals. However, adapting to the application of screening disease is challenging in that labeled ECG data are limited. Achieving general representation through self-supervised learning (SSL) is a well-known approach to overcome the scarcity of labeled data; however, a naive application of SSL to ECG data, without considering the spatial-temporal relationships inherent in ECG signals, may yield suboptimal results. In this paper, we introduce ST-MEM (Spatio-Temporal Masked Electrocardiogram Modeling), designed to learn spatio-temporal features by reconstructing masked 12-lead ECG data. ST-MEM outperforms other SSL baseline methods in various experimental settings for arrhythmia classification tasks. Moreover, we demonstrate that ST-MEM is adaptable to various lead combinations. Through quantitative and qualitative analysis, we show a spatio-temporal relationship within ECG data. Our code is available at https://github.com/bakqui/ST-MEM.
Consider a predictor, a learner, whose input is a stream of discrete items. The predictor's task, at every time point, is probabilistic multiclass prediction, i.e. to predict which item may occur next by outputting zero or more candidate items, each with a probability, after which the actual item is revealed and the predictor updates. To output probabilities, the predictor keeps track of the proportions of the items it has seen. The stream is unbounded (lifelong), and the predictor has finite limited space. The task is open-ended: the set of items is unknown to the predictor and their totality can also grow unbounded. Moreover, there is non-stationarity: the underlying frequencies of items may change, substantially, from time to time. For instance, new items may start appearing and a few recently frequent items may cease to occur again. The predictor, being space-bounded, need only provide probabilities for those items which, at the time of prediction, have sufficiently high frequency, i.e., the salient items. This problem is motivated in the setting of Prediction Games, a self-supervised learning regime where concepts serve as both the predictors and the predictands, and the set of concepts grows over time, resulting in non-stationarities as new concepts are generated and used. We design and study a number of predictors, sparse moving averages(SMAs), for the task. One SMA adapts the sparse exponentiated moving average and another is based on queuing a few counts, keeping dynamic per-item histories. Evaluating the predicted probabilities, under noise and non-stationarity, presents challenges, and we discuss and develop evaluation methods, one based on bounding log-loss. We show that a combination of ideas, supporting dynamic predictand-specific learning rates, offers advantages in terms of faster adaption to change (plasticity), while also supporting low variance (stability).
Repetitive DNA (repeats) poses significant challenges for accurate and efficient genome assembly and sequence alignment. This is particularly true for metagenomic data, where genome dynamics such as horizontal gene transfer, gene duplication, and gene loss/gain complicate accurate genome assembly from metagenomic communities. Detecting repeats is a crucial first step in overcoming these challenges. To address this issue, we propose GraSSRep, a novel approach that leverages the assembly graph's structure through graph neural networks (GNNs) within a self-supervised learning framework to classify DNA sequences into repetitive and non-repetitive categories. Specifically, we frame this problem as a node classification task within a metagenomic assembly graph. In a self-supervised fashion, we rely on a high-precision (but low-recall) heuristic to generate pseudo-labels for a small proportion of the nodes. We then use those pseudo-labels to train a GNN embedding and a random forest classifier to propagate the labels to the remaining nodes. In this way, GraSSRep combines sequencing features with pre-defined and learned graph features to achieve state-of-the-art performance in repeat detection. We evaluate our method using simulated and synthetic metagenomic datasets. The results on the simulated data highlight our GraSSRep's robustness to repeat attributes, demonstrating its effectiveness in handling the complexity of repeated sequences. Additionally, our experiments with synthetic metagenomic datasets reveal that incorporating the graph structure and the GNN enhances our detection performance. Finally, in comparative analyses, GraSSRep outperforms existing repeat detection tools with respect to precision and recall.
Exponential Moving Average (EMA) is a widely used weight averaging (WA) regularization to learn flat optima for better generalizations without extra cost in deep neural network (DNN) optimization. Despite achieving better flatness, existing WA methods might fall into worse final performances or require extra test-time computations. This work unveils the full potential of EMA with a single line of modification, i.e., switching the EMA parameters to the original model after each epoch, dubbed as Switch EMA (SEMA). From both theoretical and empirical aspects, we demonstrate that SEMA can help DNNs to reach generalization optima that better trade-off between flatness and sharpness. To verify the effectiveness of SEMA, we conduct comparison experiments with discriminative, generative, and regression tasks on vision and language datasets, including image classification, self-supervised learning, object detection and segmentation, image generation, video prediction, attribute regression, and language modeling. Comprehensive results with popular optimizers and networks show that SEMA is a free lunch for DNN training by improving performances and boosting convergence speeds.