Generative adversarial networks (GANs) have shown great success in applications such as image generation and inpainting. To stabilize the challenging training process, one typically requires large datasets - which are not available for many tasks. Large amounts of additionally available incomplete observations could be exploited in many cases, but it remains unclear how to train a GAN in such a setting. To address this shortcoming, we factorise the high-dimensional joint distribution of the complete data into a set of lower-dimensional distributions along with their dependencies. As a consequence, we can split the discriminator in a GAN into multiple "sub-discriminators" that can be independently trained from incomplete observations. Their outputs can be combined to obtain an estimate of the density ratio between the joint real and the generator distribution, which enables training the generator as in the original GAN framework. As an additional benefit, our modularisation facilitates incorporating prior knowledge into the discriminator architecture. We apply our method to image generation, image segmentation and audio source separation, and show an improved performance compared to a standard GAN when additional incomplete training examples are available.

Music has the power of inducing emotions, and human beings exploit such a phenomenon in order to empower a variety of mental states and activities, both positively and negatively. The study of emotions and music has a long and still vibrant tradition. New findings and changes of perspective in the field are not uncommon. More recent is the field investigating music emotion recognition through computational means. Music emotion recognition (MER) is an emerging and cross-disciplinary field spanning information retrieval (audio, symbolic and metadata) and machine learning, on a strong backing of music cognition (semiology of music and psychology) and music theory.

Using proof techniques involving $L^\infty$ covering numbers, we show generalisation error bounds for deep learning with two main improvements over the state of the art. First, our bounds have no explicit dependence on the number of classes except for logarithmic factors. This holds even when formulating the bounds in terms of the $L^2$-norm of the weight matrices, while previous bounds exhibit at least a square-root dependence on the number of classes in this case. Second, we adapt the Rademacher analysis of DNNs to incorporate weight sharing---a task of fundamental theoretical importance which was previously attempted only under very restrictive assumptions. In our results, each convolutional filter contributes only once to the bound, regardless of how many times it is applied. Finally we provide a few further technical improvements, including improving the width dependence from before to after pooling. We also examine our bound's behaviour on artificial data.

Several structure learning algorithms have been proposed towards discovering causal or Bayesian Network (BN) graphs, which is a particularly challenging problem in AI. The performance of these algorithms is evaluated based on the relationship the learned graph has with respect to the ground truth graph. However, there is no agreed scoring function to determine this relationship. Moreover, this paper shows that the commonly used metrics tend to be biased in favour of graphs that minimise the number of edges. The evaluation bias is inconsistent and may lead to evaluating graphs with no edges as superior to graphs with varying numbers of correct and incorrect edges; implying that graphs that minimise edges are often favoured over more complex graphs due to bias rather than overall accuracy. While graphs that are less complex are often desirable, the current metrics encourage algorithms to optimise for simplicity, and to discover graphs with a limited number of edges that do not enable full propagation of evidence. This paper proposes a Balanced Scoring Function (BSF) that eliminates this bias by adjusting the reward function based on the difficulty of discovering an edge, or no edge, proportional to their occurrence rate in the ground truth graph. The BSF score can be used in conjunction with other traditional metrics to provide an alternative and unbiased assessment about the capability of structure learning algorithms in discovering causal or BN graphs.

Understanding generalization in reinforcement learning (RL) is a significant challenge, as many common assumptions of traditional supervised learning theory do not apply. We focus on the special class of reparameterizable RL problems, where the trajectory distribution can be decomposed using the reparametrization trick. For this problem class, estimating the expected return is efficient and the trajectory can be computed deterministically given peripheral random variables, which enables us to study reparametrizable RL using supervised learning and transfer learning theory. Through these relationships, we derive guarantees on the gap between the expected and empirical return for both intrinsic and external errors, based on Rademacher complexity as well as the PAC-Bayes bound. Our bound suggests the generalization capability of reparameterizable RL is related to multiple factors including "smoothness" of the environment transition, reward and agent policy function class. We also empirically verify the relationship between the generalization gap and these factors through simulations.

Natural gradient descent, which preconditions a gradient descent update with the Fisher information matrix of the underlying statistical model, is a way to capture partial second-order information. Several highly visible works have advocated an approximation known as the empirical Fisher, drawing connections between approximate second-order methods and heuristics like Adam. We dispute this argument by showing that the empirical Fisher---unlike the Fisher---does not generally capture second-order information. We further argue that the conditions under which the empirical Fisher approaches the Fisher (and the Hessian) are unlikely to be met in practice, and that, even on simple optimization problems, the pathologies of the empirical Fisher can have undesirable effects.

The use of propensity score methods to reduce selection bias when determining causal effects is common practice for observational studies. Although such studies in econometrics, social science, and medicine often rely on sensitive data, there has been no prior work on privatising the propensity scores used to ascertain causal effects from observed data. In this paper, we demonstrate how to privatise the propensity score and quantify how the added noise for privatisation affects the propensity score as well as subsequent causal inference. We test our methods on both simulated and real-world datasets. The results are consistent with our theoretical findings that the privatisation preserves the validity of subsequent causal analysis with high probability. More importantly, our results empirically demonstrate that the proposed solutions are practical for moderately-sized datasets.

In recent years, precision agriculture that uses modern information and communication technologies is becoming very popular. Raw and semi-processed agricultural data are usually collected through various sources, such as: Internet of Thing (IoT), sensors, satellites, weather stations, robots, farm equipment, farmers and agribusinesses, etc. Besides, agricultural datasets are very large, complex, unstructured, heterogeneous, non-standardized, and inconsistent. Hence, the agricultural data mining is considered as Big Data application in terms of volume, variety, velocity and veracity. It is a key foundation to establishing a crop intelligence platform, which will enable resource efficient agronomy decision making and recommendations. In this paper, we designed and implemented a continental level agricultural data warehouse by combining Hive, MongoDB and Cassandra. Our data warehouse capabilities: (1) flexible schema; (2) data integration from real agricultural multi datasets; (3) data science and business intelligent support; (4) high performance; (5) high storage; (6) security; (7) governance and monitoring; (8) replication and recovery; (9) consistency, availability and partition tolerant; (10) distributed and cloud deployment. We also evaluate the performance of our data warehouse.

In this paper, we consider recommender systems with side information in the form of graphs. Existing collaborative filtering algorithms mainly utilize only immediate neighborhood information and have a hard time taking advantage of deeper neighborhoods beyond 1-2 hops. The main caveat of exploiting deeper graph information is the rapidly growing time and space complexity when incorporating information from these neighborhoods. In this paper, we propose using Graph DNA, a novel Deep Neighborhood Aware graph encoding algorithm, for exploiting deeper neighborhood information. DNA encoding computes approximate deep neighborhood information in linear time using Bloom filters, a space-efficient probabilistic data structure and results in a per-node encoding that is logarithmic in the number of nodes in the graph. It can be used in conjunction with both feature-based and graph-regularization-based collaborative filtering algorithms. Graph DNA has the advantages of being memory and time efficient and providing additional regularization when compared to directly using higher order graph information. We conduct experiments on real-world datasets, showing graph DNA can be easily used with 4 popular collaborative filtering algorithms and consistently leads to a performance boost with little computational and memory overhead.

Interaction graphs, such as those recording emails between individuals or transactions between institutions, tend to be sparse yet structured, and often grow in an unbounded manner. Such behavior can be well-captured by structured, nonparametric edge-exchangeable graphs. However, such exchangeable models necessarily ignore temporal dynamics in the network. We propose a dynamic nonparametric model for interaction graphs that combine the sparsity of the exchangeable models with dynamic clustering patterns that tend to reinforce recent behavioral patterns. We show that our method yields improved held-out likelihood over stationary variants, and impressive predictive performance against a range of state-of-the-art dynamic interaction graph models.