Continual Learning (CL) has achieved rapid progress in recent years. However, it is still largely unknown how to determine whether a CL model is trustworthy and how to foster its trustworthiness. This work focuses on evaluating and improving the robustness to corruptions of existing CL models. Our empirical evaluation results show that existing state-of-the-art (SOTA) CL models are particularly vulnerable to various data corruptions during testing. To make them trustworthy and robust to corruptions deployed in safety-critical scenarios, we propose a meta-learning framework of self-adaptive data augmentation to tackle the corruption robustness in CL. The proposed framework, MetaMix, learns to augment and mix data, automatically transforming the new task data or memory data. It directly optimizes the generalization performance against data corruptions during training. To evaluate the corruption robustness of our proposed approach, we construct several CL corruption datasets with different levels of severity. We perform comprehensive experiments on both task- and class-continual learning. Extensive experiments demonstrate the effectiveness of our proposed method compared to SOTA baselines.

Meta-learning has shown great promise in few-shot image classification where only a small amount of labeled data is available in each classification task. Many training tasks are provided to train a meta-model that can quickly learn new and similar concepts with few labeled samples. Data augmentation is often used to augment training tasks to avoid overfitting. However, existing data augmentation methods are often manually designed and fixed during training, ignoring training dynamics and the difference between various meta-learning settings specified by meta-model architectures and meta-learning algorithms. To address this problem, we add a task transformation layer between a training task and a meta-model such that the right amount of perturbation is added to training tasks for a certain meta-learning setting at a certain training stage. By jointly optimizing the task transformation layer and the meta-model, we avoid the risk of providing tasks that are either too easy or too difficult during training. We design the task transformation layer as a stochastic transformation function, adding the flexibility in how a training task can be transformed. We leverage differentiable data augmentations as the building blocks of the task transformation function for efficient optimization. Extensive experiments show that our method can consistently improve the few-shot generalization performance of various meta-models trained with different meta-learning algorithms, meta-model architectures, and datasets.

Continual learning (CL) aims to learn a non-stationary data distribution and not forget previous knowledge. The effectiveness of existing approaches that rely on memory replay can decrease over time as the model tends to overfit the stored examples. As a result, the model's ability to generalize well is significantly constrained. Additionally, these methods often overlook the inherent uncertainty in the memory data distribution, which differs significantly from the distribution of all previous data examples. To overcome these issues, we propose a principled memory evolution framework that dynamically adjusts the memory data distribution. This evolution is achieved by employing distributionally robust optimization (DRO) to make the memory buffer increasingly difficult to memorize. We consider two types of constraints in DRO: f-divergence and Wasserstein ball constraints. For f-divergence constraint, we derive a family of methods to evolve the memory buffer data in the continuous probability measure space with Wasserstein gradient flow (WGF). For Wasserstein ball constraint, we directly solve it in the euclidean space. Extensive experiments on existing benchmarks demonstrate the effectiveness of the proposed methods for alleviating forgetting. As a by-product of the proposed framework, our method is more robust to adversarial examples than compared CL methods.

Task-free continual learning (CL) aims to learn a non-stationary data stream without explicit task definitions and not forget previous knowledge. The widely adopted memory replay approach could gradually become less effective for long data streams, as the model may memorize the stored examples and overfit the memory buffer. Second, existing methods overlook the high uncertainty in the memory data distribution since there is a big gap between the memory data distribution and the distribution of all the previous data examples. To address these problems, for the first time, we propose a principled memory evolution framework to dynamically evolve the memory data distribution by making the memory buffer gradually harder to be memorized with distributionally robust optimization (DRO). We then derive a family of methods to evolve the memory buffer data in the continuous probability measure space with Wasserstein gradient flow (WGF). The proposed DRO is w.r.t the worst-case evolved memory data distribution, thus guarantees the model performance and learns significantly more robust features than existing memory-replay-based methods. Extensive experiments on existing benchmarks demonstrate the effectiveness of the proposed methods for alleviating forgetting. As a by-product of the proposed framework, our method is more robust to adversarial examples than existing task-free CL methods.

Existing meta-learning works assume that each task has available training and testing data. However, there are many available pre-trained models without accessing their training data in practice. We often need a single model to solve different tasks simultaneously as this is much more convenient to deploy the models. Our work aims to meta-learn a model initialization from these pre-trained models without using corresponding training data. We name this challenging problem setting as Data-Free Learning To Learn (DFL2L). We propose a distributionally robust optimization (DRO) framework to learn a black-box model to fuse and compress all the pre-trained models into a single network to address this problem. To encourage good generalization to the unseen new tasks, the proposed DRO framework diversifies the learned task embedding associated with each pre-trained model to cover the diversity in the underlying training task distributions. A model initialization is sampled from the black-box network during meta-testing as the meta learned initialization. Extensive experiments on offline and online DFL2L settings and several real image datasets demonstrate the effectiveness of the proposed methods.

Recognizing new objects by learning from a few labeled examples in an evolving environment is crucial to obtain excellent generalization ability for real-world machine learning systems. A typical setting across current meta learning algorithms assumes a stationary task distribution during meta training. In this paper, we explore a more practical and challenging setting where task distribution changes over time with domain shift. Particularly, we consider realistic scenarios where task distribution is highly imbalanced with domain labels unavailable in nature. We propose a kernel-based method for domain change detection and a difficulty-aware memory management mechanism that jointly considers the imbalanced domain size and domain importance to learn across domains continuously. Furthermore, we introduce an efficient adaptive task sampling method during meta training, which significantly reduces task gradient variance with theoretical guarantees. Finally, we propose a challenging benchmark with imbalanced domain sequences and varied domain difficulty. We have performed extensive evaluations on the proposed benchmark, demonstrating the effectiveness of our method.

Accurate traffic prediction is indispensable for intelligent traffic management. The availability of large-scale road sensing data collected by connected wireless sensors and mobile devices have provided unrealized potential for traffic prediction. However, sensory data is often incomplete due to various factors in the process of data acquisition and transmission. The missingness of traffic data brings a key challenge to the traffic prediction task since the state-of-the-art ML-based traffic prediction models (e.g., Graph Convolutional Networks (GCN)) often rely on spatial and temporal completion of the data. Moreover, existing GCN-based methods usually build a static graph based on geographical distances and are limited in their ability to capture the time-evolving relationships amongst road segments. In this paper, we develop a heterogeneous spatio-temporal prediction framework for traffic prediction using incomplete historical data. In the framework, we build multiple graphs to explicitly model the dynamic correlations among road segments from both geographical and historical aspects, and employ recurrent neural networks to capture temporal correlations for each road segment. We impute missing values in a recurrent process, which is seamlessly embedded in the prediction framework so they can be jointly trained. The proposed framework is evaluated on a public dataset of static sensors and a private dataset collected by our roving sensor system. Experimental results show the effectiveness of the proposed framework compared to state-of-the-art methods, and indicate the potential to be deployed into real-world traffic prediction systems.

Missing data, which commonly appears in multivariate time series, has been widely recognized as a key challenge in time series analysis. Many commonly used imputation methods either ignore the temporal dependencies of time series data, or do not adequately utilize the relationships among variables. State-ofthe-art methods on time series imputation are built on Recurrent Neural Networks (RNNs), which utilize the historical information to estimate current values sequentially. However, RNNs rely heavily on the output of nearby timestamps, which may lead to important information lost for long sequences. Moreover, individual variables typically present different dynamics and missingness patterns, which is neglected by the global RNN hidden states. In this paper, we propose an imputation framework to learn both global and local dependencies of multivariate time series, as well as a multi-dimensional self-attention to learn capture distant correlations across both time and feature. Extensive experiments show that the proposed framework outperforms the state-of-the-art methods in the imputation task, and benefits the downstream task.

Annotated data samples in real-world applications are often limited. Meta-learning, which utilizes prior knowledge learned from related tasks and generalizes to new tasks of limited supervised experience, is an effective approach for few-shot learning. However, standard meta-learning with globally shared knowledge cannot handle the task heterogeneity problem well, i.e., tasks lie in different distributions. Recent advances have explored several ways to trigger task-dependent initial parameters or metrics, in order to customize task-specific information. These approaches learn task contextual information from data, but ignore external domain knowledge that can help in the learning process. In this paper, we propose a task-adaptive network (TAdaNet) that makes use of a domain-knowledge graph to enrich data representations and provide task-specific customization. Specifically, we learn a task embedding that characterizes task relationships and tailors task-specific parameters, resulting in a task-adaptive metric space for classification. Experimental results on a few-shot image classification problem show the effectiveness of the proposed method. We also apply it on a real-world disease classification problem, and show promising results for clinical decision support.

The goal of metric learning is to learn a good distance metric that can capture the relationships among instances, and its importance has long been recognized in many fields. An implicit assumption in the traditional settings of metric learning is that the associated labels of the instances are deterministic. However, in many real-world applications, the associated labels come naturally with probabilities instead of deterministic values, which makes it difficult for the existing metric-learning methods to work well in these applications. To address this challenge, in this article, we study how to effectively learn the distance metric from datasets that contain probabilistic information, and then propose several novel metric-learning mechanisms for two types of probabilistic labels, i.e., the instance-wise probabilistic label and the group-wise probabilistic label. Compared with the existing metric-learning methods, our proposed mechanisms are capable of learning distance metrics directly from the probabilistic labels with high accuracy. We also theoretically analyze the proposed mechanisms and conduct extensive experiments on real-world datasets to verify the desirable properties of these mechanisms.

Utilizing multiple modalities to learn a good distance metric is of vital importance for various clinical applications. However, it is common that modalities are incomplete for some patients due to various technical and practical reasons in healthcare datasets. Existing metric learning methods cannot directly learn the distance metric on such data with missing modalities. Nevertheless, the incomplete data contains valuable information to characterize patient similarity and modality relationships, and they should not be ignored during the learning process. To tackle the aforementioned challenges, we propose a metric learning framework to perform missing modality completion and multi-modal metric learning simultaneously. Employing the generative adversarial networks, we incorporate both complete and incomplete data to learn the mapping relationship between modalities. After completing the missing modalities, we use the nonlinear representations extracted by the discriminator to learn the distance metric among patients. Through jointly training the adversarial generation part and metric learning, the similarity among patients can be learned on data with missing modalities. Experimental results show that the proposed framework learns more accurate distance metric on real-world healthcare datasets with incomplete modalities, comparing with the state-of-the-art approaches. Meanwhile, the quality of the generated modalities can be preserved.

A clinically meaningful distance metric, which is learned from measuring patient similarity, plays an important role in clinical decision support applications. Several metric learning approaches have been proposed to measure patient similarity, but they are mostly designed for learning the metric at only one time point/interval. It leads to a problem that those approaches cannot reflect the similarity variations among patients with the progression of diseases. In order to capture similarity information from multiple future time points simultaneously, we formulate a multi-task metric learning approach to identify patient similarity. However, it is challenging to directly apply traditional multi-task metric learning methods to learn such similarities due to the high dimensional, complex and noisy nature of healthcare data. Besides, the disease labels often have clinical relationships, which should not be treated as independent. Unfortunately, traditional formulation of the loss function ignores the degree of labels' similarity. To tackle the aforementioned challenges, we propose mtTSML, a multi-task triplet constrained sparse metric learning method, to monitor the similarity progression of patient pairs. In the proposed model, the distance for each task can be regarded as the combination of a common part and a task-specific one in the transformed low-rank space. We then perform sparse feature selection for each individual task to select the most discriminative information. Moreover, we use triplet constraints to guarantee the margin between similar and less similar pairs according to the ordered information of disease severity levels (i.e. labels). The experimental results on two real-world healthcare datasets show that the proposed multi-task metric learning method significantly outperforms the state-of-the-art baselines, including both single-task and multi-task metric learning methods.

Metric learning aims to learn a good distance metric that can capture the relationships among instances, and its importance has long been recognized in many fields. In the traditional settings of metric learning, an implicit assumption is that the associated labels of the instances are deterministic. However, in many real-world applications, the associated labels come naturally with probabilities instead of deterministic values. Thus, the existing metric learning methods cannot work well in these applications. To tackle this challenge, in this paper, we study how to effectively learn the distance metric from datasets that contain probabilistic information, and then propose two novel metric learning mechanisms for two types of probabilistic labels, i.e., the instance-wise probabilistic label and the group-wise probabilistic label. Compared with the existing metric learning methods, our proposed mechanisms are capable of learning distance metrics directly from the probabilistic labels with high accuracy. We also theoretically analyze the two proposed mechanisms and provide theoretical bounds on the sample complexity for both of them. Additionally, extensive experiments based on real-world datasets are conducted to verify the desirable properties of the proposed mechanisms.

Predicting the risk of potential diseases from Electronic Health Records (EHR) has attracted considerable attention in recent years, especially with the development of deep learning techniques. Compared with traditional machine learning models, deep learning based approaches achieve superior performance on risk prediction task. However, none of existing work explicitly takes prior medical knowledge (such as the relationships between diseases and corresponding risk factors) into account. In medical domain, knowledge is usually represented by discrete and arbitrary rules. Thus, how to integrate such medical rules into existing risk prediction models to improve the performance is a challenge. To tackle this challenge, we propose a novel and general framework called PRIME for risk prediction task, which can successfully incorporate discrete prior medical knowledge into all of the state-of-the-art predictive models using posterior regularization technique. Different from traditional posterior regularization, we do not need to manually set a bound for each piece of prior medical knowledge when modeling desired distribution of the target disease on patients. Moreover, the proposed PRIME can automatically learn the importance of different prior knowledge with a log-linear model.Experimental results on three real medical datasets demonstrate the effectiveness of the proposed framework for the task of risk prediction

Patient similarity learning aims to derive a clinically meaningful similarity metric to measure the similarity between a pair of patients according to their historical clinical information, which could help to predict the clinical outcomes of the patient of interest. However, the patient clinical data are usually complex, and contain much irrelevant and redundant information, which makes it difficult to learn the similarity metric with high accuracy. Although some methods have been proposed to address the complex nature of patient data, they overemphasize sparsity-based relevant feature selection and fail to take into consideration the redundant features that are highly correlated with each other, and this heavily degrades the accuracy of the learned results. To address the above challenges, we propose a novel uncorrelated patient similarity learning approach, which can not only select the most relevant features for the learning task, but also guarantee that the selected features have low correlations with each other. Additionally, to address the scenarios where the patient data are distributed across different sites, we extend the proposed approach and design a distributed mechanism, based on which the similarity metric can be accurately learned without directly accessing the raw patient data at each site. The desirable performance of the proposed methods are verified through extensive experiments conducted on both real-world and synthetic datasets.

Predicting patients' risk of developing certain diseases is an important research topic in healthcare. Personalized predictive modeling, which focuses on building specific models for individual patients, has shown its advantages on utilizing heterogeneous health data compared to global models trained on the entire population. Personalized predictive models use information from similar patient cohorts, in order to capture the specific characteristics. Accurately identifying and ranking the similarity among patients based on their historical records is a key step in personalized modeling. The electric health records (EHRs), which are irregular sampled and have varied patient visit lengths, cannot be directly used to measure patient similarity due to lack of an appropriate vector representation. In this paper, we build a novel time fusion CNN framework to simultaneously learn patient representations and measure pairwise similarity. Compared to a traditional CNN, our time fusion CNN can learn not only the local temporal relationships but also the contributions from each time interval. Along with the similarity learning process, the output information which is the probability distribution is used to rank similar patients. Utilizing the similarity scores, we perform personalized disease predictions, and compare the effect of different vector representations and similarity learning metrics.

Time series data mining has gained increasing attention in health domain. Recently, researchers attempt to employ Natural Language Processing (NLP) to health data mining, in order to learn proper representations of discrete medical concepts from Electronic Health Records (EHRs). However, existing models do not take continuous physiological records into account, which are naturally existed in EHRs. The major challenges for this task are to model non-obvious representations from observed high dimensional biosignals, and to interpret the learned features. To address these issues, we propose Wave2Vec, an end-to-end deep learning model, to bridge the gap between biosignal processing and language modeling. Wave2Vec jointly learns both inherent and embedding representations of biosignals at the same time. To evaluate the performance of our model in clinical task, we carry out experiments on two real world benchmark biosignal datasets. Experimental results show that the proposed Wave2Vec model outperforms the six feature leaning baselines in biosignal processing.