We propose a two-sample testing procedure based on learned deep neural network representations. To this end, we define two test statistics that perform an asymptotic location test on data samples mapped onto a hidden layer. The tests are consistent and asymptotically control the type-1 error rate. Their test statistics can be evaluated in linear time (in the sample size). Suitable data representations are obtained in a data-driven way, by solving a supervised or unsupervised transfer-learning task on an auxiliary (potentially distinct) data set. If no auxiliary data is available, we split the data into two chunks: one for learning representations and one for computing the test statistic. In experiments on audio samples, natural images and three-dimensional neuroimaging data our tests yield significant decreases in type-2 error rate (up to 35 percentage points) compared to state-of-the-art two-sample tests such as kernel-methods and classifier two-sample tests.

We address tracking and prediction of multiple moving objects in visual data streams as inference and sampling in a disentangled latent state-space model. By encoding objects separately and including explicit position information in the latent state space, we perform tracking via amortized variational Bayesian inference of the respective latent positions. Inference is implemented in a modular neural framework tailored towards our disentangled latent space. Generative and inference model are jointly learned from observations only. Comparing to related prior work, we empirically show that our Markovian state-space assumption enables faithful and much improved long-term prediction well beyond the training horizon. Further, our inference model correctly decomposes frames into objects, even in the presence of occlusions. Tracking performance is increased significantly over prior art.

Surrogate models such as Gaussian processes (GP) have been proposed to accelerate approximate Bayesian computation (ABC) when the statistical model of interest is expensive-to-simulate. In one such promising framework the discrepancy between simulated and observed data is modelled with a GP. So far principled strategies have been proposed only for sequential selection of the simulation locations. To address this limitation, we develop Bayesian optimal design strategies to parallellise the expensive simulations. Current surrogate-based ABC methods also produce only a point estimate of the ABC posterior while there can be substantial additional uncertainty due to the limited budget of simulations. We also address the problem of quantifying the uncertainty of ABC posterior and discuss the connections between our resulting framework called Bayesian ABC, Bayesian quadrature (BQ) and Bayesian optimisation (BO). Experiments with several toy and real-world simulation models demonstrate advantages of the proposed techniques.

Sleep staging is a crucial task for diagnosing sleep disorders. It is tedious and complex as it can take a trained expert several hours to annotate just one patient's polysomnogram (PSG) from a single night. Although deep learning models have demonstrated state-of-the-art performance in automating sleep staging, interpretability which defines other desiderata, has largely remained unexplored. In this study, we propose Sleep staging via Prototypes from Expert Rules (SLEEPER), which combines deep learning models with expert defined rules using a prototype learning framework to generate simple interpretable models. In particular, SLEEPER utilizes sleep scoring rules and expert defined features to derive prototypes which are embeddings of PSG data fragments via convolutional neural networks. The final models are simple interpretable models like a shallow decision tree defined over those phenotypes. We evaluated SLEEPER using two PSG datasets collected from sleep studies and demonstrated that SLEEPER could provide accurate sleep stage classification comparable to human experts and deep neural networks with about 85% ROC-AUC and .7 kappa.

We propose a new algorithm for adversarial multi-armed bandits with unrestricted delays. The algorithm is based on a novel hybrid regularizer applied in the Follow the Regularized Leader (FTRL) framework. It achieves $\mathcal{O}(\sqrt{kn}+\sqrt{D\log(k)})$ regret guarantee, where $k$ is the number of arms, $n$ is the number of rounds, and $D$ is the total delay. The result matches the lower bound within constants and requires no prior knowledge of $n$ or $D$. Additionally, we propose a refined tuning of the algorithm, which achieves $\mathcal{O}(\sqrt{kn}+\min_{S}|S|+\sqrt{D_{\bar S}\log(k)})$ regret guarantee, where $S$ is a set of rounds excluded from delay counting, $\bar S = [n]\setminus S$ are the counted rounds, and $D_{\bar S}$ is the total delay in the counted rounds. If the delays are highly unbalanced, the latter regret guarantee can be significantly tighter than the former. The result requires no advance knowledge of the delays and resolves an open problem of Thune et al. (2019). The new FTRL algorithm and its refined tuning are anytime and require no doubling, which resolves another open problem of Thune et al. (2019).

To coordinate actions with an interaction partner requires a constant exchange of sensorimotor signals. Humans acquire these skills in infancy and early childhood mostly by imitation learning and active engagement with a skilled partner. They require the ability to predict and adapt to one's partner during an interaction. In this work we want to explore these ideas in a human-robot interaction setting in which a robot is required to learn interactive tasks from a combination of observational and kinesthetic learning. To this end, we propose a deep learning framework consisting of a number of components for (1) human and robot motion embedding, (2) motion prediction of the human partner and (3) generation of robot joint trajectories matching the human motion. To test these ideas, we collect human-human interaction data and human-robot interaction data of four interactive tasks "hand-shake", "hand-wave", "parachute fist-bump" and "rocket fist-bump". We demonstrate experimentally the importance of predictive and adaptive components as well as low-level abstractions to successfully learn to imitate human behavior in interactive social tasks.

We study the problem of recovering clusters from binary user feedback. Items are grouped into initially unknown non-overlapping clusters. To recover these clusters, the learner sequentially presents to users a finite list of items together with a question with a binary answer selected from a fixed finite set. For each of these items, the user provides a random answer whose expectation is determined by the item cluster and the question and by an item-specific parameter characterizing the hardness of classifying the item. The objective is to devise an algorithm with a minimal cluster recovery error rate. We derive problem-specific information-theoretical lower bounds on the error rate satisfied by any algorithm, for both uniform and adaptive (list, question) selection strategies. For uniform selection, we present a simple algorithm built upon K-means whose performance almost matches the fundamental limits. For adaptive selection, we develop an adaptive algorithm that is inspired by the derivation of the information-theoretical error lower bounds, and in turn allocates the budget in an efficient way. The algorithm learns to select items hard to cluster and relevant questions more often. We compare numerically the performance of our algorithms with or without adaptive selection strategy, and illustrate the gain achieved by being adaptive. Our inference problems are motivated by the problem of solving large-scale labeling tasks with minimal effort put on the users. For example, in some of the recent CAPTCHA systems, users clicks (binary answers) can be used to efficiently label images, by optimally finding the best questions to present.

The Fisher information matrix (FIM) plays an essential role in statistics and machine learning as a Riemannian metric tensor. Focusing on the FIM and its variants in deep neural networks (DNNs), we reveal their characteristic behavior when the network is sufficiently wide and has random weights and biases. Various FIMs asymptotically show pathological eigenvalue spectra in the sense that a small number of eigenvalues take on large values while most of them are close to zero. This implies that the local shape of the parameter space or loss landscape is very steep in a few specific directions and almost flat in the other directions. Similar pathological spectra appear in other variants of FIMs: one is the neural tangent kernel; another is a metric for the input signal and feature space that arises from feedforward signal propagation. The quantitative understanding of the FIM and its variants provided here offers important perspectives on learning and signal processing in large-scale DNNs.

The breakthrough of deep Q-Learning on different types of environments revolutionized the algorithmic design of Reinforcement Learning to introduce more stable and robust algorithms, to that end many extensions to deep Q-Learning algorithm have been proposed to reduce the variance of the target values and the overestimation phenomena. In this paper, we examine new methodology to solve these issues, we propose using Dropout techniques on deep Q-Learning algorithm as a way to reduce variance and overestimation. We further present experiments on some of the benchmark environments that demonstrate significant improvement of the stability of the performance and a reduction in variance and overestimation.

We consider the kernel completion problem with the presence of multiple views in the data. In this context the data samples can be fully missing in some views, creating missing columns and rows to the kernel matrices that are calculated individually for each view. We propose to solve the problem of completing the kernel matrices by transferring the features of the other views to represent the view under consideration. We align the known part of the kernel matrix with a new kernel built from the features of the other views. We are thus able to find generalizable structures in the kernel under completion, and represent it accurately. Its missing values can be predicted with the data available in other views. We illustrate the benefits of our approach with simulated data and multivariate digits dataset, as well as with real biological datasets from studies of pattern formation in early \textit{Drosophila melanogaster} embryogenesis.