This paper presents a novel neural network training approach for faster convergence and better generalization abilities in deep reinforcement learning. Particularly, we focus on the enhancement of training and evaluation performance in reinforcement learning algorithms by systematically reducing gradient's variance and thereby providing a more targeted learning process. The proposed method which we term as Gradient Monitoring(GM), is an approach to steer the learning in the weight parameters of a neural network based on the dynamic development and feedback from the training process itself. We propose different variants of the GM methodology which have been proven to increase the underlying performance of the model. The one of the proposed variant, Momentum with Gradient Monitoring (M-WGM), allows for a continuous adjustment of the quantum of back-propagated gradients in the network based on certain learning parameters. We further enhance the method with Adaptive Momentum with Gradient Monitoring (AM-WGM) method which allows for automatic adjustment between focused learning of certain weights versus a more dispersed learning depending on the feedback from the rewards collected. As a by-product, it also allows for automatic derivation of the required deep network sizes during training as the algorithm automatically freezes trained weights. The approach is applied to two discrete (Multi-Robot Co-ordination problem and Atari games) and one continuous control task (MuJoCo) using Advantage Actor-Critic (A2C) and Proximal Policy Optimization (PPO) respectively. The results obtained particularly underline the applicability and performance improvements of the methods in terms of generalization capability.

Clinical neuroimaging has recently witnessed explosive growth in data availability which brings studying heterogeneity in clinical cohorts to the spotlight. Normative modeling is an emerging statistical tool for achieving this objective. However, its application remains technically challenging due to difficulties in properly dealing with nuisance variation, for example due to variability in image acquisition devices. Here, in a fully probabilistic framework, we propose an application of hierarchical Bayesian regression (HBR) for multi-site normative modeling. Our experimental results confirm the superiority of HBR in deriving more accurate normative ranges on large multi-site neuroimaging data compared to widely used methods. This provides the possibility i) to learn the normative range of structural and functional brain measures on large multi-site data; ii) to recalibrate and reuse the learned model on local small data; therefore, HBR closes the technical loop for applying normative modeling as a medical tool for the diagnosis and prognosis of mental disorders.

We study anomaly detection and introduce an algorithm that processes variable length, irregularly sampled sequences or sequences with missing values. Our algorithm is fully unsupervised, however, can be readily extended to supervised or semisupervised cases when the anomaly labels are present as remarked throughout the paper. Our approach uses the Long Short Term Memory (LSTM) networks in order to extract temporal features and find the most relevant feature vectors for anomaly detection. We incorporate the sampling time information to our model by modulating the standard LSTM model with time modulation gates. After obtaining the most relevant features from the LSTM, we label the sequences using a Support Vector Data Descriptor (SVDD) model. We introduce a loss function and then jointly optimize the feature extraction and sequence processing mechanisms in an end-to-end manner. Through this joint optimization, the LSTM extracts the most relevant features for anomaly detection later to be used in the SVDD, hence completely removes the need for feature selection by expert knowledge. Furthermore, we provide a training algorithm for the online setup, where we optimize our model parameters with individual sequences as the new data arrives. Finally, on real-life datasets, we show that our model significantly outperforms the standard approaches thanks to its combination of LSTM with SVDD and joint optimization.

We construct deep ReLU neural networks to approximate functions in dilated shift-invariant spaces generated by a continuous function with compact support and study the approximation rates with respect to the number of neurons. The network construction is based on the bit extraction and data fitting capacity of deep neural networks. Combining with existing results of approximation from shift-invariant spaces, we are able to estimate the approximation rates of classical function spaces such as Sobolev spaces and Besov spaces. We also give lower bounds of the $L^p([0,1]^d)$ approximation error for Sobolev spaces, which show that our construction is asymptotically optimal up to a logarithm factor.

Land cover maps are a vital input variable to many types of environmental research and management. While they can be produced automatically by machine learning techniques, these techniques require substantial training data to achieve high levels of accuracy, which are not always available. One technique researchers use when labelled training data are scarce is domain adaptation (DA) -- where data from an alternate region, known as the source domain, are used to train a classifier and this model is adapted to map the study region, or target domain. The scenario we address in this paper is known as semi-supervised DA, where some labelled samples are available in the target domain. In this paper we present Sourcerer, a Bayesian-inspired, deep learning-based, semi-supervised DA technique for producing land cover maps from SITS data. The technique takes a convolutional neural network trained on a source domain and then trains further on the available target domain with a novel regularizer applied to the model weights. The regularizer adjusts the degree to which the model is modified to fit the target data, limiting the degree of change when the target data are few in number and increasing it as target data quantity increases. Our experiments on Sentinel-2 time series images compare Sourcerer with two state-of-the-art semi-supervised domain adaptation techniques and four baseline models. We show that on two different source-target domain pairings Sourcerer outperforms all other methods for any quantity of labelled target data available. In fact, the results on the more difficult target domain show that the starting accuracy of Sourcerer (when no labelled target data are available), 74.2%, is greater than the next-best state-of-the-art method trained on 20,000 labelled target instances.

Semi-supervised learning algorithms attempt to take advantage of relatively inexpensive unlabeled data to improve learning performance. In this work, we consider statistical models where the data distributions can be characterized by continuous parameters. We show that under certain conditions on the distribution, unlabeled data is equally useful as labeled date in terms of learning rate. Specifically, let $n, m$ be the number of labeled and unlabeled data, respectively. It is shown that the learning rate of semi-supervised learning scales as $O(1/n)$ if $m\sim n$, and scales as $O(1/n^{1+\gamma})$ if $m\sim n^{1+\gamma}$ for some $\gamma>0$, whereas the learning rate of supervised learning scales as $O(1/n)$.

Multiclass classification problems are most often solved by either training a single centralized classifier on all $K$ classes, or by reducing the problem to multiple binary classification tasks. This paper explores the uncharted region between these two extremes: How can we solve the $K$-class classification problem by combining the predictions of smaller classifiers, each trained on an arbitrary number of classes $R \in \{2, 3, \ldots, K\}$? We present a mathematical framework for answering this question, and derive bounds on the number of classifiers (in terms of $K$ and $R$) needed to accurately predict the true class of an unlabeled sample under both adversarial and stochastic assumptions. By exploiting a connection to the classical set cover problem in combinatorics, we produce an efficient, near-optimal scheme (with respect to the number of classifiers) for designing such configurations of classifiers, which recovers the well-known one-vs.-one strategy as a special case when $R=2$. Experiments with the MNIST and CIFAR-10 datasets show that our scheme is capable of matching the performance of centralized classifiers in practice. The results suggest that our approach offers a promising direction for solving the problem of data heterogeneity which plagues current federated learning methods.

We study statistical and computational limits of clustering when the means of the centres are sparse and their dimension is possibly much larger than the sample size. Our theoretical analysis focuses on the simple model $X_i = z_i \theta + \varepsilon_i$, $z_i \in \{-1,1\}$, $\varepsilon_i \thicksim \mathcal{N}(0, I)$, which has two clusters with centres $\theta$ and $-\theta$. We provide a finite sample analysis of a new sparse clustering algorithm based on sparse PCA and show that it achieves the minimax optimal misclustering rate in the regime $\|\theta\| \rightarrow \infty$, matching asymptotically the Bayes error. Our results require the sparsity to grow slower than the square root of the sample size. Using a recent framework for computational lower bounds---the low-degree likelihood ratio---we give evidence that this condition is necessary for any polynomial-time clustering algorithm to succeed below the BBP threshold. This complements existing evidence based on reductions and statistical query lower bounds. Compared to these existing results, we cover a wider set of parameter regimes and give a more precise understanding of the runtime required and the misclustering error achievable. We also discuss extensions of our results to more than two clusters.

Detecting anomalies in dynamic graphs is a vital task, with numerous practical applications in areas such as security, finance, and social media. Previous network embedding based methods have been mostly focusing on learning good node representations, whereas largely ignoring the subgraph structural changes related to the target nodes in dynamic graphs. In this paper, we propose StrGNN, an end-to-end structural temporal Graph Neural Network model for detecting anomalous edges in dynamic graphs. In particular, we first extract the $h$-hop enclosing subgraph centered on the target edge and propose the node labeling function to identify the role of each node in the subgraph. Then, we leverage graph convolution operation and Sortpooling layer to extract the fixed-size feature from each snapshot/timestamp. Based on the extracted features, we utilize Gated recurrent units (GRUs) to capture the temporal information for anomaly detection. Extensive experiments on six benchmark datasets and a real enterprise security system demonstrate the effectiveness of StrGNN.

In the domain of social signal processing, audio recognition is a promising avenue for accessing daily behaviors that contribute to health and well-being. However, despite advances in mobile computing and machine learning, audio behavior detection models are largely constrained to data collected in controlled settings, such as call centers. This is problematic as it means their performance is unlikely to generalize to real-world applications. In the current paper, we present a model combining deep spectrum and acoustic features to detect and classify infant distress vocalizations from 24 hour, continuous, raw real-world data collected via a wearable audio recorder. Our model dramatically outperforms infant distress detection models trained and tested on equivalent real-world datasets. In particular, our model has an F1 score of 0.597 relative to F1 scores of 0.166 and 0.26 achieved by state-of-practice and state-of-the-art real-world infant distress classifiers, respectively. We end by discussing what may have facilitated this massive gain in accuracy, including using supervised deep spectrum features and the fact that we collected and annotated a massive dataset of 780 hours of real-world audio data with over 25 hours of labelled distress.