Sleep stage classification is critical for diagnosing sleep quality. While deep neural networks are becoming popular for automatic sleep stage classification with supervised learning, large-scale labeled datasets are still hard to acquire. Recently, self-supervised learning (SSL) has become one of the most prominent approaches to alleviate the burden of labeling works. However, existing SSL methods are mainly designed for non-temporally correlated data and only learning representations from an instance level. Hence, the objective of this paper is to learn robust and generalizable representations for physiological signals with self-supervised learning. Specifically, we make the following contributions: (1) we propose a novel co-training scheme by exploiting complementary information from multiple views (time view and frequency view) of physiological signals to mine more positive samples, which overcomes the drawback of popular InfoNCE and achieves semantic-level representation learning; (2) we extend our framework with a memory module, implemented by a queue and a moving-averaged encoder, to enlarge the pool of negative candidates and keep the up-to-date representation; (3) extensive experiments conducted on sleep stage classification demonstrate state-of-the-art performance compared with SSL baselines, achieving 71.6% and 57.9% accuracies on two sleep datasets, SleepEDF and ISRUC respectively. The code is publicly available at https://github.com/larryshaw0079/CoSleep .

We address the problem of unsupervised anomaly detection for multivariate data. Traditional machine learning based anomaly detection algorithms rely on specific assumptions of normal patterns and fail to model complex feature interactions and relations. Recently, existing deep learning based methods are promising for extracting representations from complex features. These methods train an auxiliary task, e.g., reconstruction and prediction, on normal samples. They further assume that anomalies fail to perform well on the auxiliary task since they are never trained during the model optimization. However, the assumption does not always hold in practice. Deep models may also perform the auxiliary task well on anomalous samples, leading to the failure detection of anomalies. To effectively detect anomalies for multivariate data, this paper introduces a teacher-student distillation based framework Distillated Teacher-Student Network Ensemble (DTSNE). The paradigm of the teacher-student distillation is able to deal with high-dimensional complex features. In addition, an ensemble of student networks provides a better capability to avoid generalizing the auxiliary task performance on anomalous samples. To validate the effectiveness of our model, we conduct extensive experiments on real-world datasets. Experimental results show superior performance of DTSNE over competing methods. Analysis and discussion towards the behavior of our model are also provided in the experiment section.

Optimizing spaced repetition schedules is of great importance for enhancing long-term memory retention in both real-world applications, e.g., online learning platforms, and academic applications, e.g., cognitive science. Traditional methods tackle this problem by employing handcrafted rules while modern methods try to optimize scheduling using deep reinforcement learning (DRL). Existing DRL-based approaches model the problem by selecting the optimal next item to appear, which implies the learner can only learn one item in a day. However, the most essential point to enhancing long-term memory is to select the optimal interval to review. To this end, we present a novel approach to DRL to optimize spaced repetition scheduling. The contribution of our framework is three-fold. We first introduce a Transformer-based model to estimate the recall probability of a learning item accurately, which encodes the temporal dynamics of a learner’s learning trajectories. Second, we build a simulation environment based on our recall probability estimation model. Third, we utilize the Deep Q-Network (DQN) as the agent to learn the optimal review intervals for learning items and train the policy in a recurrent manner. Experimental results demonstrate that our framework achieves state-of-the-art performance against competing methods. Our method achieves an MAE (mean average error) score of $0.0274$ on a memory prediction task, which is $11\%$ lower than the second-best method. For spaced repetition scheduling, our method achieves mean recall probabilities of $0.92$, $0.942$, and $0.372$ in three different environments, the best performance in all scenarios.