Conventionally, Earth system (e.g., weather and climate) forecasting relies on numerical simulation with complex physical models and hence is both expensive in computation and demanding on domain expertise. With the explosive growth of spatiotemporal Earth observation data in the past decade, data-driven models that apply Deep Learning (DL) are demonstrating impressive potential for various Earth system forecasting tasks. The Transformer as an emerging DL architecture, despite its broad success in other domains, has limited adoption in this area. In this paper, we propose Earthformer, a space-time Transformer for Earth system forecasting. Earthformer is based on a generic, flexible and efficient space-time attention block, named Cuboid Attention. The idea is to decompose the data into cuboids and apply cuboid-level self-attention in parallel. These cuboids are further connected with a collection of global vectors. We conduct experiments on the MovingMNIST dataset and a newly proposed chaotic N-body MNIST dataset to verify the effectiveness of cuboid attention and figure out the best design of Earthformer. Experiments on two real-world benchmarks about precipitation nowcasting and El Niño/Southern Oscillation (ENSO) forecasting show that Earthformer achieves state-of-the-art performance.

In this work, we propose a deep learning approach to forecasting state-level COVID-19 trends of weekly cumulative death in the United States (US) and incident cases in Germany. This approach includes a transformer model, an ensemble method, and a data augmentation technique for time series. We arrange the inputs of the transformer in such a way that predictions for different states can attend to the trends of the others. To overcome the issue of scarcity of training data for this COVID-19 pandemic, we have developed a novel data augmentation technique to generate useful data for training. More importantly, the generated data can also be used for model validation. As such, it has a two-fold advantage: 1) more actual observations can be used for training, and 2) the model can be validated on data which has distribution closer to the expected situation. Our model has achieved some of the best state-level results on the COVID-19 Forecast Hub for the US and for Germany.

The effectiveness of learning in massive open online courses (MOOCs) can be significantly enhanced by introducing personalized intervention schemes which rely on building predictive models of student learning behaviors such as some engagement or performance indicators. A major challenge that has to be addressed when building such models is to design handcrafted features that are effective for the prediction task at hand. In this paper, we make the first attempt to solve the feature learning problem by taking the unsupervised learning approach to learn a compact representation of the raw features with a large degree of redundancy. Specifically, in order to capture the underlying learning patterns in the content domain and the temporal nature of the clickstream data, we train a modified auto-encoder (AE) combined with the long short-term memory (LSTM) network to obtain a fixed-length embedding for each input sequence. When compared with the original features, the new features that correspond to the embedding obtained by the modified LSTM-AE are not only more parsimonious but also more discriminative for our prediction task. Using simple supervised learning models, the learned features can improve the prediction accuracy by up to 17% compared with the supervised neural networks and reduce overfitting to the dominant low-performing group of students, specifically in the task of predicting students' performance. Our approach is generic in the sense that it is not restricted to a specific supervised learning model nor a specific prediction task for MOOC learning analytics.

Spatiotemporal systems are common in the real-world. Forecasting the multi-step future of these spatiotemporal systems based on the past observations, or, Spatiotemporal Sequence Forecasting (STSF), is a significant and challenging problem. Although lots of real-world problems can be viewed as STSF and many research works have proposed machine learning based methods for them, no existing work has summarized and compared these methods from a unified perspective. This survey aims to provide a systematic review of machine learning for STSF. In this survey, we define the STSF problem and classify it into three subcategories: Trajectory Forecasting of Moving Point Cloud (TF-MPC), STSF on Regular Grid (STSF-RG) and STSF on Irregular Grid (STSF-IG). We then introduce the two major challenges of STSF: 1) how to learn a model for multi-step forecasting and 2) how to adequately model the spatial and temporal structures. After that, we review the existing works for solving these challenges, including the general learning strategies for multi-step forecasting, the classical machine learning based methods for STSF, and the deep learning based methods for STSF. We also compare these methods and point out some potential research directions.

Knowledge tracing is one of the key research areas for empowering personalized education. It is a task to model students' mastery level of a knowledge component (KC) based on their historical learning trajectories. In recent years, a recurrent neural network model called deep knowledge tracing (DKT) has been proposed to handle the knowledge tracing task and literature has shown that DKT generally outperforms traditional methods. However, through our extensive experimentation, we have noticed two major problems in the DKT model. The first problem is that the model fails to reconstruct the observed input. As a result, even when a student performs well on a KC, the prediction of that KC's mastery level decreases instead, and vice versa. Second, the predicted performance for KCs across time-steps is not consistent. This is undesirable and unreasonable because student's performance is expected to transit gradually over time. To address these problems, we introduce regularization terms that correspond to reconstruction and waviness to the loss function of the original DKT model to enhance the consistency in prediction. Experiments show that the regularized loss function effectively alleviates the two problems without degrading the original task of DKT.

With the goal of making high-resolution forecasts of regional rainfall, precipitation nowcasting has become an important and fundamental technology underlying various public services ranging from rainstorm warnings to flight safety. Recently, the Convolutional LSTM (ConvLSTM) model has been shown to outperform traditional optical flow based methods for precipitation nowcasting, suggesting that deep learning models have a huge potential for solving the problem. However, the convolutional recurrence structure in ConvLSTM-based models is location-invariant while natural motion and transformation (e.g., rotation) are location-variant in general. Furthermore, since deep-learning-based precipitation nowcasting is a newly emerging area, clear evaluation protocols have not yet been established. To address these problems, we propose both a new model and a benchmark for precipitation nowcasting. Specifically, we go beyond ConvLSTM and propose the Trajectory GRU (TrajGRU) model that can actively learn the location-variant structure for recurrent connections. Besides, we provide a benchmark that includes a real-world large-scale dataset from the Hong Kong Observatory, a new training loss, and a comprehensive evaluation protocol to facilitate future research and gauge the state of the art.

Many problems in image processing and computer vision (e.g. colorization, style transfer) can be posed as 'manipulating' an input image into a corresponding output image given a user-specified guiding signal. A holy-grail solution towards generic image manipulation should be able to efficiently alter an input image with any personalized signals (even signals unseen during training), such as diverse paintings and arbitrary descriptive attributes. However, existing methods are either inefficient to simultaneously process multiple signals (let alone generalize to unseen signals), or unable to handle signals from other modalities. In this paper, we make the first attempt to address the zero-shot image manipulation task. We cast this problem as manipulating an input image according to a parametric model whose key parameters can be conditionally generated from any guiding signal (even unseen ones). To this end, we propose the Zero-shot Manipulation Net (ZM-Net), a fully-differentiable architecture that jointly optimizes an image-transformation network (TNet) and a parameter network (PNet). The PNet learns to generate key transformation parameters for the TNet given any guiding signal while the TNet performs fast zero-shot image manipulation according to both signal-dependent parameters from the PNet and signal-invariant parameters from the TNet itself. Extensive experiments show that our ZM-Net can perform high-quality image manipulation conditioned on different forms of guiding signals (e.g. style images and attributes) in real-time (tens of milliseconds per image) even for unseen signals. Moreover, a large-scale style dataset with over 20,000 style images is also constructed to promote further research.

Despite recent advances in the visual tracking community, most studies so far have focused on the observation model. As another important component in the tracking system, the motion model is much less well-explored especially for some extreme scenarios. In this paper, we consider one such scenario in which the camera is mounted on an unmanned aerial vehicle (UAV) or drone. We build a benchmark dataset of high diversity, consisting of 70 videos captured by drone cameras. To address the challenging issue of severe camera motion, we devise simple baselines to model the camera motion by geometric transformation based on background feature points. An extensive comparison of recent state-of-the-art trackers and their motion model variants on our drone tracking dataset validates both the necessity of the dataset and the effectiveness of the proposed methods. Our aim for this work is to lay the foundation for further research in the UAV tracking area.

We are interested in exploring the possibility and benefits of structure learning for deep models. As the first step, this paper investigates the matter for Restricted Boltzmann Machines (RBMs) . We conduct the study with Replicated Softmax, a variant of RBMs for unsupervised text analysis. We present a method for learning what we call Sparse Boltzmann Machines , where each hidden unit is connected to a subset of the visible units instead of all of them.  Empirical results show that the method yields models with significantly improved model fit and interpretability as compared with RBMs where each hidden unit is connected to all visible units.

Link prediction is a fundamental task in such areas as social network analysis, information retrieval, and bioinformatics. Usually link prediction methods use the link structures or node attributes as the sources of information. Recently, the relational topic model (RTM) and its variants have been proposed as hybrid methods that jointly model both sources of information and achieve very promising accuracy. However, the representations (features) learned by them are still not effective enough to represent the nodes (items). To address this problem, we generalize recent advances in deep learning from solely modeling i.i.d. sequences of attributes to jointly modeling graphs and non-i.i.d. sequences of attributes. Specifically, we follow the Bayesian deep learning framework and devise a hierarchical Bayesian model, called relational deep learning (RDL), to jointly model high-dimensional node attributes and link structures with layers of latent variables. Due to the multiple nonlinear transformations in RDL, standard variational inference is not applicable. We propose to utilize the product of Gaussians (PoG) structure in RDL to relate the inferences on different variables and derive a generalized variational inference algorithm for learning the variables and predicting the links. Experiments on three real-world datasets show that RDL works surprisingly well and significantly outperforms the state of the art.