Generative modeling and self-supervised learning have in recent years made great strides towards learning from data in a completely unsupervised way. There is still however an open area of investigation into guiding a neural network to encode the data into representations that are interpretable or explainable. The problem of unsupervised disentanglement is of particular importance as it proposes to discover the different latent factors of variation or semantic concepts from the data alone, without labeled examples, and encode them into structurally disjoint latent representations. Without additional constraints or inductive biases placed in the network, a generative model may learn the data distribution and encode the factors, but not necessarily in a disentangled way. Here, we introduce a novel deep generative VAE-based model, ProtoVAE, that leverages a deep metric learning Prototypical network trained using self-supervision to impose these constraints. The prototypical network constrains the mapping of the representation space to data space to ensure that controlled changes in the representation space are mapped to changes in the factors of variations in the data space. Our model is completely unsupervised and requires no a priori knowledge of the dataset, including the number of factors. We evaluate our proposed model on the benchmark dSprites, 3DShapes, and MPI3D disentanglement datasets, showing state of the art results against previous methods via qualitative traversals in the latent space, as well as quantitative disentanglement metrics. We further qualitatively demonstrate the effectiveness of our model on the real-world CelebA dataset.
Self-supervised learning (SSL) speech models such as wav2vec and HuBERT have demonstrated state-of-the-art performance on automatic speech recognition (ASR) and proved to be extremely useful in low label-resource settings. However, the success of SSL models has yet to transfer to utterance-level tasks such as speaker, emotion, and language recognition, which still require supervised fine-tuning of the SSL models to obtain good performance. We argue that the problem is caused by the lack of disentangled representations and an utterance-level learning objective for these tasks. Inspired by how HuBERT uses clustering to discover hidden acoustic units, we formulate a factor analysis (FA) model that uses the discovered hidden acoustic units to align the SSL features. The underlying utterance-level representations are disentangled from the content of speech using probabilistic inference on the aligned features. Furthermore, the variational lower bound derived from the FA model provides an utterance-level objective, allowing error gradients to be backpropagated to the Transformer layers to learn highly discriminative acoustic units. When used in conjunction with HuBERT's masked prediction training, our models outperform the current best model, WavLM, on all utterance-level non-semantic tasks on the SUPERB benchmark with only 20% of labeled data.
Designing an effective representation learning method for multimodal sentiment analysis tasks is a crucial research direction. The challenge lies in learning both shared and private information in a complete modal representation, which is difficult with uniform multimodal labels and a raw feature fusion approach. In this work, we propose a deep modal shared information learning module based on the covariance matrix to capture the shared information between modalities. Additionally, we use a label generation module based on a self-supervised learning strategy to capture the private information of the modalities. Our module is plug-and-play in multimodal tasks, and by changing the parameterization, it can adjust the information exchange relationship between the modes and learn the private or shared information between the specified modes. We also employ a multi-task learning strategy to help the model focus its attention on the modal differentiation training data. We provide a detailed formulation derivation and feasibility proof for the design of the deep modal shared information learning module. We conduct extensive experiments on three common multimodal sentiment analysis baseline datasets, and the experimental results validate the reliability of our model. Furthermore, we explore more combinatorial techniques for the use of the module. Our approach outperforms current state-of-the-art methods on most of the metrics of the three public datasets.
Self-supervised learning (SSL) enables label efficient training for machine learning models. This is essential for domains such as medical imaging, where labels are costly and time-consuming to curate. However, the most effective supervised or SSL strategy for transferring models to different healthcare systems or novel tasks is not well understood. In this work, we systematically experiment with a variety of supervised and self-supervised pretraining strategies using multimodal datasets of medical images (chest X-rays) and text (radiology reports). We then evaluate their performance on data from two external institutions with diverse sets of tasks. In addition, we experiment with different transfer learning strategies to effectively adapt these pretrained models to new tasks and healthcare systems. Our empirical results suggest that multimodal SSL gives substantial gains over unimodal SSL in performance across new healthcare systems and tasks, comparable to models pretrained with full supervision. We demonstrate additional performance gains with models further adapted to the new dataset and task, using multimodal domain-adaptive pretraining (DAPT), linear probing then finetuning (LP-FT), and both methods combined. We offer suggestions for alternative models to use in scenarios where not all of these additions are feasible. Our results provide guidance for improving the generalization of medical image interpretation models to new healthcare systems and novel tasks.
This paper tries to address a fundamental question in point cloud self-supervised learning: what is a good signal we should leverage to learn features from point clouds without annotations? To answer that, we introduce a point cloud representation learning framework, based on geometric feature reconstruction. In contrast to recent papers that directly adopt masked autoencoder (MAE) and only predict original coordinates or occupancy from masked point clouds, our method revisits differences between images and point clouds and identifies three self-supervised learning objectives peculiar to point clouds, namely centroid prediction, normal estimation, and curvature prediction. Combined with occupancy prediction, these four objectives yield an nontrivial self-supervised learning task and mutually facilitate models to better reason fine-grained geometry of point clouds. Our pipeline is conceptually simple and it consists of two major steps: first, it randomly masks out groups of points, followed by a Transformer-based point cloud encoder; second, a lightweight Transformer decoder predicts centroid, normal, and curvature for points in each voxel. We transfer the pre-trained Transformer encoder to a downstream peception model. On the nuScene Datset, our model achieves 3.38 mAP improvment for object detection, 2.1 mIoU gain for segmentation, and 1.7 AMOTA gain for multi-object tracking. We also conduct experiments on the Waymo Open Dataset and achieve significant performance improvements over baselines as well.
Contrastive learning has emerged as an efficient framework to learn multimodal representations. CLIP, a seminal work in this area, achieved impressive results by training on paired image-text data using the contrastive loss. Recent work claims improvements over CLIP using additional non-contrastive losses inspired from self-supervised learning. However, it is sometimes hard to disentangle the contribution of these additional losses from other implementation details, e.g., data augmentation or regularization techniques, used to train the model. To shed light on this matter, in this paper, we first propose, implement and evaluate several baselines obtained by combining contrastive learning with recent advances in self-supervised learning. In particular, we use the loss functions that were proven successful for visual self-supervised learning to align image and text modalities. We find that these baselines outperform a basic implementation of CLIP. However, when a stronger training recipe is employed, the advantage disappears. Indeed, we find that a simple CLIP baseline can also be improved substantially, up to a 25% relative improvement on downstream zero-shot tasks, by using well-known training techniques that are popular in other subfields. Moreover, we discover that it is enough to apply image and text augmentations to make up for most of the improvement attained by prior works. With our improved training recipe for CLIP, we obtain state-of-the-art performance on four standard datasets, and consistently outperform prior work (up to +4% on the largest dataset), while being substantially simpler. The code is available at https://github.com/facebookresearch/clip-rocket
Natural environments such as forests and grasslands are challenging for robotic navigation because of the false perception of rigid obstacles from high grass, twigs, or bushes. In this work, we propose Wild Visual Navigation (WVN), an online self-supervised learning system for traversability estimation which uses only vision. The system is able to continuously adapt from a short human demonstration in the field. It leverages high-dimensional features from self-supervised visual transformer models, with an online scheme for supervision generation that runs in real-time on the robot. We demonstrate the advantages of our approach with experiments and ablation studies in challenging environments in forests, parks, and grasslands. Our system is able to bootstrap the traversable terrain segmentation in less than 5 min of in-field training time, enabling the robot to navigate in complex outdoor terrains - negotiating obstacles in high grass as well as a 1.4 km footpath following. While our experiments were executed with a quadruped robot, ANYmal, the approach presented can generalize to any ground robot.
Large-scale multi-modal contrastive learning frameworks like CLIP typically require a large amount of image-text samples for training. However, these samples are always collected continuously in real scenarios. This paper discusses the feasibility of continual CLIP training using streaming data. Unlike continual learning based on self-supervised learning methods for pure images, which is empirically robust against catastrophic forgetting, CLIP's performance degeneration in the continual setting is significant and non-neglectable. By analyzing the changes in the model's representation space during continual CLIP training from a spatial geometry perspective, we explore and summarize these spatial variations as Spatial Disorder (SD), which can be divided into Intra-modal Rotation and Inter-modal Deviation. Moreover, we empirically and theoretically demonstrate how SD leads to a performance decline for CLIP on cross-modal retrieval tasks. To alleviate SD, we propose a new continual vision-language representation learning framework Mod-X: Maintain off-diagonal information-matriX. By selectively aligning the off-diagonal information distribution of contrastive matrices, the Mod-X improves the capability of the multi-modal model by maintaining the multi-modal representation space alignment on the old data domain during continuously fitting the new training data domain. Experiments on commonly used datasets with different scales and scopes have demonstrated the effectiveness of our method.
The success of self-supervised learning in computer vision and natural language processing has motivated pretraining methods on tabular data. However, most existing tabular self-supervised learning models fail to leverage information across multiple data tables and cannot generalize to new tables. In this work, we introduce XTab, a framework for cross-table pretraining of tabular transformers on datasets from various domains. We address the challenge of inconsistent column types and quantities among tables by utilizing independent featurizers and using federated learning to pretrain the shared component. Tested on 84 tabular prediction tasks from the OpenML-AutoML Benchmark (AMLB), we show that (1) XTab consistently boosts the generalizability, learning speed, and performance of multiple tabular transformers, (2) by pretraining FT-Transformer via XTab, we achieve superior performance than other state-of-the-art tabular deep learning models on various tasks such as regression, binary, and multiclass classification.
Semantic localization, i.e., robot self-localization with semantic image modality, is critical in recently emerging embodied AI applications (e.g., point-goal navigation, object-goal navigation, vision language navigation) and topological mapping applications (e.g., graph neural SLAM, ego-centric topological map). However, most existing works on semantic localization focus on passive vision tasks without viewpoint planning, or rely on additional rich modalities (e.g., depth measurements). Thus, the problem is largely unsolved. In this work, we explore a lightweight, entirely CPU-based, domain-adaptive semantic localization framework, called graph neural localizer. Our approach is inspired by two recently emerging technologies: (1) Scene graph, which combines the viewpoint- and appearance- invariance of local and global features; (2) Graph neural network, which enables direct learning/recognition of graph data (i.e., non-vector data). Specifically, a graph convolutional neural network is first trained as a scene graph classifier for passive vision, and then its knowledge is transferred to a reinforcement-learning planner for active vision. Experiments on two scenarios, self-supervised learning and unsupervised domain adaptation, using a photo-realistic Habitat simulator validate the effectiveness of the proposed method.
Self-supervised learning (SSL) has been dramatically successful not only in monolingual but also in cross-lingual settings. However, since the two settings have been studied individually in general, there has been little research focusing on how effective a cross-lingual model is in comparison with a monolingual model. In this paper, we investigate this fundamental question empirically with Japanese automatic speech recognition (ASR) tasks. First, we begin by comparing the ASR performance of cross-lingual and monolingual models for two different language tasks while keeping the acoustic domain as identical as possible. Then, we examine how much unlabeled data collected in Japanese is needed to achieve performance comparable to a cross-lingual model pre-trained with tens of thousands of hours of English and/or multilingual data. Finally, we extensively investigate the effectiveness of SSL in Japanese and demonstrate state-of-the-art performance on multiple ASR tasks. Since there is no comprehensive SSL study for Japanese, we hope this study will guide Japanese SSL research.
Recent advances in self-supervised learning and neural network scaling have enabled the creation of large models, known as foundation models, which can be easily adapted to a wide range of downstream tasks. The current paradigm for comparing foundation models involves evaluating them with aggregate metrics on various benchmark datasets. This method of model comparison is heavily dependent on the chosen evaluation metric, which makes it unsuitable for situations where the ideal metric is either not obvious or unavailable. In this work, we present a methodology for directly comparing the embedding space geometry of foundation models, which facilitates model comparison without the need for an explicit evaluation metric. Our methodology is grounded in random graph theory and enables valid hypothesis testing of embedding similarity on a per-datum basis. Further, we demonstrate how our methodology can be extended to facilitate population level model comparison. In particular, we show how our framework can induce a manifold of models equipped with a distance function that correlates strongly with several downstream metrics. We remark on the utility of this population level model comparison as a first step towards a taxonomic science of foundation models.
The search for emotional biomarkers within the human voice is a challenging research area. Previous studies focused on predicting affective state from speech; this study explores various tasks on affective vocal bursts. Borrowing the success of self-supervised learning in automatic speech recognition, we extracted acoustic embedding using variants of wav2vec 2.0 for four affective vocal bursts tasks: High, Two, Culture, and Type. Using a similar architecture for all tasks, the evaluation of acoustic embeddings reveals the potential use of wav2vec 2.0 variants over the conventional acoustic features in affective vocal bursts tasks. We evaluated both conventional acoustic features and these acoustic embeddings on the different number of twenty seeds evaluation and reported the maximum and average scores with their standard deviations in the validation set. Three high scores from these validations for all tasks assist the generation of predictions for the test set. We compared the test scores with previous studies and obtained remarkable improvements.
Self-supervised methods based on contrastive learning have achieved great success in unsupervised visual representation learning. However, most methods under this framework suffer from the problem of false negative samples. Inspired by the mean shift for self-supervised learning, we propose a new simple framework, namely Multiple Sample Views and Queues (MSVQ). We jointly construct three soft labels on-the-fly by utilizing two complementary and symmetric approaches: multiple augmented positive views and two momentum encoders that generate various semantic features for negative samples. Two teacher networks perform similarity relationship calculations with negative samples and then transfer this knowledge to the student network. Let the student network mimic the similarity relationships between the samples, thus giving the student network a more flexible ability to identify false negative samples in the dataset. The classification results on four benchmark image datasets demonstrate the high effectiveness and efficiency of our approach compared to some classical methods. Source code and pretrained models are available \href{https://github.com/pc-cp/MSVQ}{here}.
Over the last few years, massive amounts of satellite multispectral and hyperspectral images covering the Earth's surface have been made publicly available for scientific purpose, for example through the European Copernicus project. Simultaneously, the development of self-supervised learning (SSL) methods has sparked great interest in the remote sensing community, enabling to learn latent representations from unlabeled data to help treating downstream tasks for which there is few annotated examples, such as interpolation, forecasting or unmixing. Following this line, we train a deep learning model inspired from the Koopman operator theory to model long-term reflectance dynamics in an unsupervised way. We show that this trained model, being differentiable, can be used as a prior for data assimilation in a straightforward way. Our datasets, which are composed of Sentinel-2 multispectral image time series, are publicly released with several levels of treatment.
Self-supervised learning can extract representations of good quality from solely unlabeled data, which is appealing for point cloud videos due to their high labelling cost. In this paper, we propose a contrastive mask prediction (PointCMP) framework for self-supervised learning on point cloud videos. Specifically, our PointCMP employs a two-branch structure to achieve simultaneous learning of both local and global spatio-temporal information. On top of this two-branch structure, a mutual similarity based augmentation module is developed to synthesize hard samples at the feature level. By masking dominant tokens and erasing principal channels, we generate hard samples to facilitate learning representations with better discrimination and generalization performance. Extensive experiments show that our PointCMP achieves the state-of-the-art performance on benchmark datasets and outperforms existing full-supervised counterparts. Transfer learning results demonstrate the superiority of the learned representations across different datasets and tasks.
Hand gesture serves as a crucial role during the expression of sign language. Current deep learning based methods for sign language understanding (SLU) are prone to over-fitting due to insufficient sign data resource and suffer limited interpretability. In this paper, we propose the first self-supervised pre-trainable SignBERT+ framework with model-aware hand prior incorporated. In our framework, the hand pose is regarded as a visual token, which is derived from an off-the-shelf detector. Each visual token is embedded with gesture state and spatial-temporal position encoding. To take full advantage of current sign data resource, we first perform self-supervised learning to model its statistics. To this end, we design multi-level masked modeling strategies (joint, frame and clip) to mimic common failure detection cases. Jointly with these masked modeling strategies, we incorporate model-aware hand prior to better capture hierarchical context over the sequence. After the pre-training, we carefully design simple yet effective prediction heads for downstream tasks. To validate the effectiveness of our framework, we perform extensive experiments on three main SLU tasks, involving isolated and continuous sign language recognition (SLR), and sign language translation (SLT). Experimental results demonstrate the effectiveness of our method, achieving new state-of-the-art performance with a notable gain.
Point cloud data has been extensively studied due to its compact form and flexibility in representing complex 3D structures. The ability of point cloud data to accurately capture and represent intricate 3D geometry makes it an ideal choice for a wide range of applications, including computer vision, robotics, and autonomous driving, all of which require an understanding of the underlying spatial structures. Given the challenges associated with annotating large-scale point clouds, self-supervised point cloud representation learning has attracted increasing attention in recent years. This approach aims to learn generic and useful point cloud representations from unlabeled data, circumventing the need for extensive manual annotations. In this paper, we present a comprehensive survey of self-supervised point cloud representation learning using DNNs. We begin by presenting the motivation and general trends in recent research. We then briefly introduce the commonly used datasets and evaluation metrics. Following that, we delve into an extensive exploration of self-supervised point cloud representation learning methods based on these techniques. Finally, we share our thoughts on some of the challenges and potential issues that future research in self-supervised learning for pre-training 3D point clouds may encounter.
Change detection from traditional \added{2D} optical images has limited capability to model the changes in the height or shape of objects. Change detection using 3D point cloud \added{from photogrammetry or LiDAR surveying} can fill this gap by providing critical depth information. While most existing machine learning based 3D point cloud change detection methods are supervised, they severely depend on the availability of annotated training data, which is in practice a critical point. To circumnavigate this dependence, we propose an unsupervised 3D point cloud change detection method mainly based on self-supervised learning using deep clustering and contrastive learning. The proposed method also relies on an adaptation of deep change vector analysis to 3D point cloud via nearest point comparison. Experiments conducted on \added{an aerial LiDAR survey dataset} show that the proposed method obtains higher performance in comparison to the traditional unsupervised methods, with a gain of about 9\% in mean accuracy (to reach more than 85\%). Thus, it appears to be a relevant choice in scenario where prior knowledge (labels) is not ensured.
Humans can learn incrementally, whereas neural networks forget previously acquired information catastrophically. Continual Learning (CL) approaches seek to bridge this gap by facilitating the transfer of knowledge to both previous tasks (backward transfer) and future ones (forward transfer) during training. Recent research has shown that self-supervision can produce versatile models that can generalize well to diverse downstream tasks. However, contrastive self-supervised learning (CSSL), a popular self-supervision technique, has limited effectiveness in online CL (OCL). OCL only permits one iteration of the input dataset, and CSSL's low sample efficiency hinders its use on the input data-stream. In this work, we propose Continual Learning via Equivariant Regularization (CLER), an OCL approach that leverages equivariant tasks for self-supervision, avoiding CSSL's limitations. Our method represents the first attempt at combining equivariant knowledge with CL and can be easily integrated with existing OCL methods. Extensive ablations shed light on how equivariant pretext tasks affect the network's information flow and its impact on CL dynamics.