One of the central questions in statistical learning theory is to determine the conditions under which agents can learn from experience. This includes the necessary and sufficient conditions for generalization from a given finite training set to new observations. In this paper, we prove that algorithmic stability in the inference process is equivalent to uniform generalization across all parametric loss functions. We provide various interpretations of this result. For instance, a relationship is proved between stability and data processing, which reveals that algorithmic stability can be improved by post-processing the inferred hypothesis or by augmenting training examples with artificial noise prior to learning.

We propose a reinforcement learning agent to solve hard exploration games by learning a range of directed exploratory policies. We construct an episodic memory-based intrinsic reward using k-nearest neighbors over the agent's recent experience to train the directed exploratory policies, thereby encouraging the agent to repeatedly revisit all states in its environment. A self-supervised inverse dynamics model is used to train the embeddings of the nearest neighbour lookup, biasing the novelty signal towards what the agent can control. We employ the framework of Universal Value Function Approximators (UVFA) to simultaneously learn many directed exploration policies with the same neural network, with different trade-offs between exploration and exploitation. By using the same neural network for different degrees of exploration/exploitation, transfer is demonstrated from predominantly exploratory policies yielding effective exploitative policies. The proposed method can be incorporated to run with modern distributed RL agents that collect large amounts of experience from many actors running in parallel on separate environment instances. Our method doubles the performance of the base agent in all hard exploration in the Atari-57 suite while maintaining a very high score across the remaining games, obtaining a median human normalised score of 1344.0%. Notably, the proposed method is the first algorithm to achieve non-zero rewards (with a mean score of 8,400) in the game of Pitfall! without using demonstrations or hand-crafted features.

In many applications, accurate class probability estimates are required, but many types of models produce poor quality probability estimates despite achieving acceptable classification accuracy. Even though probability calibration has been a hot topic of research in recent times, the majority of this has investigated non-sequential data. In this paper, we consider calibrating models that produce class probability estimates from sequences of data, focusing on the case where predictions are obtained from incomplete sequences. We show that traditional calibration techniques are not sufficiently expressive for this task, and propose methods that adapt calibration schemes depending on the length of an input sequence. Experimental evaluation shows that the proposed methods are often substantially more effective at calibrating probability estimates from modern sequential architectures for incomplete sequences across a range of application domains.

For organizing large text corpora topic modeling provides useful tools. A widely used method is Latent Dirichlet Allocation (LDA), a generative probabilistic model which models single texts in a collection of texts as mixtures of latent topics. The assignments of words to topics rely on initial values such that generally the outcome of LDA is not fully reproducible. In addition, the reassignment via Gibbs Sampling is based on conditional distributions, leading to different results in replicated runs on the same text data. This fact is often neglected in everyday practice. We aim to improve the reliability of LDA results. Therefore, we study the stability of LDA by comparing assignments from replicated runs. We propose to quantify the similarity of two generated topics by a modified Jaccard coefficient. Using such similarities, topics can be clustered. A new pruning algorithm for hierarchical clustering results based on the idea that two LDA runs create pairs of similar topics is proposed. This approach leads to the new measure S-CLOP ({\bf S}imilarity of multiple sets by {\bf C}lustering with {\bf LO}cal {\bf P}runing) for quantifying the stability of LDA models. We discuss some characteristics of this measure and illustrate it with an application to real data consisting of newspaper articles from \textit{USA Today}. Our results show that the measure S-CLOP is useful for assessing the stability of LDA models or any other topic modeling procedure that characterize its topics by word distributions. Based on the newly proposed measure for LDA stability, we propose a method to increase the reliability and hence to improve the reproducibility of empirical findings based on topic modeling. This increase in reliability is obtained by running the LDA several times and taking as prototype the most representative run, that is the LDA run with highest average similarity to all other runs.

Single-cell RNA sequencing (scRNA-seq) has revolutionized biological discovery, providing an unbiased picture of cellular heterogeneity in tissues. While scRNA-seq has been used extensively to provide insight into health and disease, it has not been used for disease prediction or diagnostics. Graph Attention Networks have proven to be versatile for a wide range of tasks by learning from both original features and graph structures. Here we present a graph attention model for predicting disease state from single-cell data on a large dataset of Multiple Sclerosis (MS) patients. MS is a disease of the central nervous system that is difficult to diagnose. We train our model on single-cell data obtained from blood and cerebrospinal fluid (CSF) for a cohort of seven MS patients and six healthy adults (HA), resulting in 66,667 individual cells. We achieve $\mathbf{92}$ \% accuracy in predicting MS, outperforming other state-of-the-art methods such as a graph convolutional network, random forest, and multi-layer perceptron. Further, we use the learned graph attention model to get insight into the features (cell types and genes) that are important for this prediction. The graph attention model also allow us to infer a new feature space for the cells that emphasizes the difference between the two conditions. Finally we use the attention weights to learn a new low-dimensional embedding which we visualize with PHATE and UMAP. To the best of our knowledge, this is the first effort to use graph attention, and deep learning in general, to predict disease state from single-cell data. We envision applying this method to single-cell data for other diseases.

An important challenge in the field of exponential random graphs (ERGs) is the fitting of non-trivial ERGs on large networks. By utilizing matrix block-approximation techniques, we propose an approximative framework to such non-trivial ERGs that result in dyadic independence (i.e., edge independent) models, while being able to meaningfully model local information (degrees) as well as global information (clustering coefficient, assortativity, etc.) if desired. This allows one to efficiently generate random networks with similar properties as an observed network, scalable up to sparse graphs consisting of millions of nodes. Empirical evaluation demonstrates its competitiveness in terms of accuracy with state-of-the-art methods for link prediction and network reconstruction.

Training a classifier over a large number of classes, known as 'extreme classification', has become a topic of major interest with applications in technology, science, and e-commerce. Traditional softmax regression induces a gradient cost proportional to the number of classes $C$, which often is prohibitively expensive. A popular scalable softmax approximation relies on uniform negative sampling, which suffers from slow convergence due a poor signal-to-noise ratio. In this paper, we propose a simple training method for drastically enhancing the gradient signal by drawing negative samples from an adversarial model that mimics the data distribution. Our contributions are three-fold: (i) an adversarial sampling mechanism that produces negative samples at a cost only logarithmic in $C$, thus still resulting in cheap gradient updates; (ii) a mathematical proof that this adversarial sampling minimizes the gradient variance while any bias due to non-uniform sampling can be removed; (iii) experimental results on large scale data sets that show a reduction of the training time by an order of magnitude relative to several competitive baselines.

A popular trick for computing a pairwise co-occurrence matrix is the product of an incidence matrix and its transpose. We present an analog for higher order tuple co-occurrences using the face-splitting product, or alternately known as the transpose Khatri-Rao product. These higher order co-occurrences encode the commonality of tokens in the company of other tokens, and thus generalize the mutual information commonly studied. We demonstrate this tensor's use via a popular NLP model, and hypergraph models of similarity. Studying an implicit meaning of a collection of things via their relationships with other things of the same type is a popular technique.

Wind power as a renewable source of energy, has numerous economic, environmental and social benefits. In order to enhance and control renewable wind power, it is vital to utilize models that predict wind speed with high accuracy. Due to neglecting of requirement and significance of data preprocessing and disregarding the inadequacy of using a single predicting model, many traditional models have poor performance in wind speed prediction. In the current study, for predicting wind speed at target stations in the north of Iran, the combination of a multi-layer perceptron model (MLP) with the Whale Optimization Algorithm (WOA) used to build new method (MLP-WOA) with a limited set of data (2004-2014). Then, the MLP-WOA model was utilized at each of the ten target stations, with the nine stations for training and tenth station for testing (namely: Astara, Bandar-E-Anzali, Rasht, Manjil, Jirandeh, Talesh, Kiyashahr, Lahijan, Masuleh, and Deylaman) to increase the accuracy of the subsequent hybrid model. The capability of the hybrid model in wind speed forecasting at each target station was compared with the MLP model without the WOA optimizer. To determine definite results, numerous statistical performances were utilized. For all ten target stations, the MLP-WOA model had precise outcomes than the standalone MLP model. The hybrid model had acceptable performances with lower amounts of the RMSE, SI and RE parameters and higher values of NSE, WI, and KGE parameters. It was concluded that the WOA optimization algorithm can improve the prediction accuracy of MLP model and may be recommended for accurate wind speed prediction.

We introduce a simple (one line of code) modification to the Generative Adversarial Network (GAN) training algorithm that materially improves results with no increase in computational cost: When updating the generator parameters, we simply zero out the gradient contributions from the elements of the batch that the critic scores as `least realistic'. Through experiments on many different GAN variants, we show that this `top-k update' procedure is a generally applicable improvement. In order to understand the nature of the improvement, we conduct extensive analysis on a simple mixture-of-Gaussians dataset and discover several interesting phenomena. Among these is that, when gradient updates are computed using the worst-scoring batch elements, samples can actually be pushed further away from the their nearest mode.