Our understanding of the world depends highly on our capacity to produce intuitive and simplified representations which can be easily used to solve problems. We reproduce this simplification process using a neural network to build a low dimensional state representation of the world from images acquired by a robot. As in Jonschkowski et al. 2015, we learn in an unsupervised way using prior knowledge about the world as loss functions called robotic priors and extend this approach to high dimension richer images to learn a 3D representation of the hand position of a robot from RGB images. We propose a quantitative evaluation of the learned representation using nearest neighbors in the state space that allows to assess its quality and show both the potential and limitations of robotic priors in realistic environments. We augment image size, add distractors and domain randomization, all crucial components to achieve transfer learning to real robots. Finally, we also contribute a new prior to improve the robustness of the representation. The applications of such low dimensional state representation range from easing reinforcement learning (RL) and knowledge transfer across tasks, to facilitating learning from raw data with more efficient and compact high level representations. The results show that the robotic prior approach is able to extract high level representation as the 3D position of an arm and organize it into a compact and coherent space of states in a challenging dataset.

Connected vehicle technology is foreseen to play an important role in reducing the number of traffic accidents while being one of the main enabling components for autonomous driving. One of the application of such connection is to provide accurate information about the road condition such as friction level to drivers or the intelligent systems controlling the car. Road surface friction can be defined as the grip between car tyre and underlying surface. During winter times when the temperature decreases dramatically, friction level reduces substantially, which can increase the risk of car accidents. Studies indicate that road conditions such as surface temperature, type of road, and structure of the road sides play an important role in the measured friction level, and some of these conditions can vary significantly within short distances under specific weather situations. Road friction prediction based on the past sensor measurements available in the cars, e.g., temperature and sun light, has advantages of being independent of the road structure and surrounding infrastructure. Intelligent forecast systems rely on the availability of high quality data in order to allow their multiple actors to make correct decisions in diverse traffic situations. These systems have the potential to increase the safety of roads users by means of the timely sharing of road-related information. With the advances in car-to-car communication technology, today, Volvo cars are equipped with slippery road condition warning system to improve road safety and traffic flow.

Abstract--This paper is concerned with estimation and stochastic control in physical systems which contain unknown input signals or forces. These unknown signals are modeled as Gaussian processes (GP) in the sense that GP models are used in machine learning. The resulting latent force models (LFMs) can be seen as hybrid models that contain a first-principles physical model part and a nonparametric GP model part. The aim of this paper is to collect and extend the statistical inference and learning methods for this kind of models, provide new theoretical results for the models, and to extend the methodology and theory to stochastic control of LFMs. The generalizations of this kind of models to arbitrary differential equations are called latent force models (LFM) [2]-[6] in machine learning literature. In addition to learning problem on the LFMs, we also consider the problem of controlling the LFM using the control functionc(t) . In particular, we consider the problem of optimal stochastic control design for LFMs. The present problem is also closely related to so called input estimation problem that has previously been addressed in target tracking literature (e.g. Simo S arkk a is with the Department of Electrical Engineering and Automation (EEA), Aalto University, Rakentajanaukio 2c, 02150 Espoo, Finland (simo.sarkka@aalto.fi). The difference is that here is no concept of time in this equation, nor a possibility for controlling the equation. A. General problem formulation The models considered in this article can be seen to belong to the following three classes: 1) Basic latent force models which are ordinary differential equations (ODEs) driven by Gaussian input processes u (t) and control inputsc(t) . X, MONTH 20XX 2 2) We also consider are dynamic partial and pseudo differential equation (PDE) based models that can generally be written in form L f (x,t) u (x,t) c(x,t), (7) where L is a linear operator in space and time. The input Gaussian processu (x,t) and control inputc(x,t) are also space-time processes. Typically, the operator has the form L A m d m dt m ··· A 1 d dt A 0, (8) where A 0,...,A m are some spatial partial differential or pseudo-differential operators. This kind of models can often be also written in form of spatiotemporal state-space models f (x,t) t A f f (x,t) B f u (x,t) M f c (x,t), (9) which again is strictly more general than the model (8). For this kind of models there is no control problem per se, because there is no time dependence. These models do not naturally allow for a state-space representation either.

It becomes increasingly popular to perform mediation analysis for complex data from sophisticated experimental studies. In this paper, we present Granger Mediation Analysis (GMA), a new framework for causal mediation analysis of multiple time series. This framework is motivated by a functional magnetic resonance imaging (fMRI) experiment where we are interested in estimating the mediation effects between a randomized stimulus time series and brain activity time series from two brain regions. The stable unit treatment assumption for causal mediation analysis is thus unrealistic for this type of time series data. To address this challenge, our framework integrates two types of models: causal mediation analysis across the variables and vector autoregressive models across the temporal observations. We further extend this framework to handle multilevel data to address individual variability and correlated errors between the mediator and the outcome variables. These models not only provide valid causal mediation for time series data but also model the causal dynamics across time. We show that the modeling parameters in our models are identifiable, and we develop computationally efficient methods to maximize the likelihood-based optimization criteria. Simulation studies show that our method reduces the estimation bias and improve statistical power, compared to existing approaches. On a real fMRI data set, our approach not only infers the causal effects of brain pathways but accurately captures the feedback effect of the outcome region on the mediator region.

We compare two statistical models of three binary random variables. One is a mixture model and the other is a product of mixtures model called a restricted Boltzmann machine. Although the two models we study look different from their parametrizations, we show that they represent the same set of distributions on the interior of the probability simplex, and are equal up to closure. We give a semi-algebraic description of the model in terms of six binomial inequalities and obtain closed form expressions for the maximum likelihood estimates. We briefly discuss extensions to larger models.

In this work, we develop a novel framework for activity shaping under the Continuous-Time Information Cascades Model which allows the administrator for local control actions by allocating targeted resources that can alter the spread of the process. Our framework employs the optimization of the spectral radius of the Hazard matrix, a quantity that has been shown to drive the maximum influence in a network, while enjoying a simple convex relaxation when used to minimize the influence of the cascade. In addition, use-cases such as quarantine and node immunization are discussed to highlight the generality of the proposed activity shaping framework. Finally, we present the NetShape influence minimization method which is compared favorably to baseline and state-of-the-art approaches through simulations on real social networks.

To model high dimensional data, Gaussian methods are widely used since they remain tractable and yield parsimonious models by imposing strong assumptions on the data. Vine copulas are more flexible by combining arbitrary marginal distributions and (conditional) bivariate copulas. Yet, this adaptability is accompanied by sharply increasing computational effort as the dimension increases. The approach proposed in this paper overcomes this burden and makes the first step into ultra high dimensional non-Gaussian dependence modeling by using a divide-and-conquer approach. First, we apply Gaussian methods to split datasets into feasibly small subsets and second, apply parsimonious and flexible vine copulas thereon. Finally, we reconcile them into one joint model. We provide numerical results demonstrating the feasibility of our approach in moderate dimensions and showcase its ability to estimate ultra high dimensional non-Gaussian dependence models in thousands of dimensions.

Predicating macroscopic influences of drugs on human body, like efficacy and toxicity, is a central problem of small-molecule based drug discovery. Molecules can be represented as an undirected graph, and we can utilize graph convolution networks to predication molecular properties. However, graph convolutional networks and other graph neural networks all focus on learning node-level representation rather than graph-level representation. Previous works simply sum all feature vectors for all nodes in the graph to obtain the graph feature vector for drug predication. In this paper, we introduce a dummy super node that is connected with all nodes in the graph by a directed edge as the representation of the graph and modify the graph operation to help the dummy super node learn graph-level feature. Thus, we can handle graph-level classification and regression in the same way as node-level classification and regression. In addition, we apply focal loss to address class imbalance in drug datasets. The experiments on MoleculeNet show that our method can effectively improve the performance of molecular properties predication.

We present Fashion-MNIST, a new dataset comprising of 28x28 grayscale images of 70,000 fashion products from 10 categories, with 7,000 images per category. The training set has 60,000 images and the test set has 10,000 images. Fashion-MNIST is intended to serve as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms, as it shares the same image size, data format and the structure of training and testing splits. The dataset is freely available at https://github.com/zalandoresearch/fashion-mnist

Most popular word embedding techniques involve implicit or explicit factorization of a word co-occurrence based matrix into low rank factors. In this paper, we aim to generalize this trend by using numerical methods to factor higher-order word co-occurrence based arrays, or \textit{tensors}. We present four word embeddings using tensor factorization and analyze their advantages and disadvantages. One of our main contributions is a novel joint symmetric tensor factorization technique related to the idea of coupled tensor factorization. We show that embeddings based on tensor factorization can be used to discern the various meanings of polysemous words without being explicitly trained to do so, and motivate the intuition behind why this works in a way that doesn't with existing methods. We also modify an existing word embedding evaluation metric known as Outlier Detection [Camacho-Collados and Navigli, 2016] to evaluate the quality of the order-$N$ relations that a word embedding captures, and show that tensor-based methods outperform existing matrix-based methods at this task. Experimentally, we show that all of our word embeddings either outperform or are competitive with state-of-the-art baselines commonly used today on a variety of recent datasets. Suggested applications of tensor factorization-based word embeddings are given, and all source code and pre-trained vectors are publicly available online.