The first food product invented by an artificial intelligence (AI) system can already be obtained in Argentina. And with this specific meal, AI shows that it can already revolutionize a productive item that, until now, was conservative: the food industry. Dictating an email to the smartphone, driverless cars or combat drones that choose a target and fire without a commander's order are already common. The novelty is that, in laboratories and companies, computational techniques of Machine Learning and Big Data are already used to create new -and recreate old- foods using absolutely novel ingredients. For example, a Chilean startup has just presented in the local market a mayonnaise that has the same taste, smell, color and texture as the traditional one.

We present a framework for recovering/approximating unknown time-dependent partial differential equation (PDE) using its solution data. Instead of identifying the terms in the underlying PDE, we seek to approximate the evolution operator of the underlying PDE numerically. The evolution operator of the PDE, defined in infinite-dimensional space, maps the solution from a current time to a future time and completely characterizes the solution evolution of the underlying unknown PDE. Our recovery strategy relies on approximation of the evolution operator in a properly defined modal space, i.e., generalized Fourier space, in order to reduce the problem to finite dimensions. The finite dimensional approximation is then accomplished by training a deep neural network structure, which is based on residual network (ResNet), using the given data. Error analysis is provided to illustrate the predictive accuracy of the proposed method. A set of examples of different types of PDEs, including inviscid Burgers' equation that develops discontinuity in its solution, are presented to demonstrate the effectiveness of the proposed method.

V ARIBAD: A V ERY G OOD M ETHOD FOR B AYES-A DAPTIVE D EEP RL VIA M ETA-L EARNING Luisa Zintgraf University of Oxford Kyriacos Shiarlis Latent Logic Maximilian Igl University of Oxford Sebastian Schulze University of Oxford Y arin Gal OA TML Group, University of Oxford Katja Hofmann Microsoft Research Shimon Whiteson University of Oxford Latent Logic A BSTRACT Trading off exploration and exploitation in an unknown environment is key to maximising expected return during learning. A Bayes-optimal policy, which does so optimally, conditions its actions not only on the environment state but on the agent's uncertainty about the environment. Computing a Bayes-optimal policy is however intractable for all but the smallest tasks. In this paper, we introduce variational Bayes-Adaptive Deep RL (variBAD), a way to meta-learn to perform approximate inference in an unknown environment, and incorporate task uncertainty directly during action selection. In a grid-world domain, we illustrate how variBAD performs structured online exploration as a function of task uncertainty. We also evaluate variBAD on MuJoCo domains widely used in meta-RL and show that it achieves higher return during training than existing methods. 1 I NTRODUCTION Reinforcement learning (RL) is typically concerned with finding an optimal policy that maximises expected return for a given Markov decision process (MDP) with an unknown reward and transition function. If these were known, the optimal policy could in theory be computed without interacting with the environment. By contrast, learning in an unknown environment typically requires trading off exploration (learning about the environment) and exploitation (taking promising actions). Balancing this tradeoff is key to maximising expected return during learning . A Bayes-optimal policy, which does so optimally, conditions actions not only on the environment state but on the agent's own uncertainty about the current MDP . In principle, a Bayes-optimal policy can be computed using the framework of Bayes-adaptive Markov decision processes (BAMDPs) (Martin, 1967; Duff & Barto, 2002). The agent maintains a belief, i.e., a posterior distribution, over possible environments. Augmenting the state space of the underlying MDP with this posterior distribution yields a BAMDP, a special case of a belief MDP (Kaelbling et al., 1998).

--Recently, randomly mapping vectorial data to strings of discrete symbols (i.e., sketches) for fast and space-efficient similarity searches has become popular . Such random mapping is called similarity-preserving hashing and approximates a similarity metric by using the Hamming distance. Although many efficient similarity searches have been proposed, most of them are designed for binary sketches. Similarity searches on integer sketches are in their infancy. In this paper, we present a novel space-efficient trie named b -bit sketch trie on integer sketches for scalable similarity searches by leveraging the idea behind succinct data structures (i.e., space-efficient data structures while supporting various data operations in the compressed format) and a favorable property of integer sketches as fixed-length strings. Our experimental results obtained using real-world datasets show that a trie-based index is built from integer sketches and efficiently performs similarity searches on the index by pruning useless portions of the search space, which greatly improves the search time and space-efficiency of the similarity search. The experimental results show that our similarity search is at most one order of magnitude faster than state-of-the-art similarity searches. Besides, our method needs only 10 GiB of memory on a billion-scale database, while state-of-the-art similarity searches need 29 GiB of memory. I NTRODUCTION The similarity search of vectorial data in databases has been a fundamental task in recent data analysis, and it has various applications such as near duplicate detection in a collection of web pages [1], context-based retrieval in images [2], and functional analysis of molecules [3]. Recently, databases in these applications have become large, and vectorial data in these databases also have been high dimensional, which makes it difficult to apply existing similarity search methods to such large databases. There is thus a strong need to develop much more powerful methods of similarity search for efficiently analyzing databases on a large-scale. A powerful solution to address this need is similarity-preserving hashing, which intends to approximate a similarity measure by randomly mapping vectorial data in a metric space to strings of discrete symbols (i.e., sketches) in the Hamming space. Early methods include Sim-Hash for cosine similarity [4], which intends to build binary sketches from vectorial data for approximating cosine similarity.

This paper is concerned with multi-view reinforcement learning (MVRL), which allows for decision making when agents share common dynamics but adhere to different observation models. We define the MVRL framework by extending partially observable Markov decision processes (POMDPs) to support more than one observation model and propose two solution methods through observation augmentation and cross-view policy transfer. We empirically evaluate our method and demonstrate its effectiveness in a variety of environments. Specifically, we show reductions in sample complexities and computational time for acquiring policies that handle multi-view environments.

A composite neural network is a rooted directed acyclic graph combining a set of pre-trained and non-instantiated neural network models. A pre-trained neural network model is well-crafted for a specific task and with instantiated weights. is generally well trained, targeted to approximate a specific function. Despite a general belief that a composite neural network may perform better than a single component, the overall performance characteristics are not clear. In this work, we prove that there exist parameters such that a composite neural network performs better than any of its pre-trained components with a high probability bound.

Understanding local currents in the North Atlantic region of the ocean is a key part of modelling heat transfer and global climate patterns. Satellites provide a surface signature of the temperature of the ocean with a high horizontal resolution while in situ autonomous probes supply high vertical resolution, but horizontally sparse, knowledge of the ocean interior thermal structure. The objective of this paper is to develop a methodology to combine these complementary ocean observing systems measurements to obtain a three-dimensional time series of ocean temperatures with high horizontal and vertical resolution. Within an observation-driven framework, we investigate the extent to which mesoscale ocean dynamics in the North Atlantic region may be decomposed into a mixture of dynamical modes, characterized by different local regressions between Sea Surface Temperature (SST), Sea Level Anomalies (SLA) and Vertical Temperature fields. Ultimately we propose a Latent-class regression method to improve prediction of vertical ocean temperature.

We present two instances, L-GAE and L-VGAE, of the variational graph auto-encoding family (VGAE) based on separating feature propagation operations from graph convolution layers typically found in graph learning methods to a single linear matrix computation made prior to input in standard auto-encoder architectures. This decoupling enables the independent and fixed design of the auto-encoder without requiring additional GCN layers for every desired increase in the size of a node's local receptive field. Fixing the auto-encoder enables a fairer assessment on the size of a nodes receptive field in building representations. Furthermore a by-product of fixing the auto-encoder design often results in substantially smaller networks than their VGAE counterparts especially as we increase the number of feature propagations. A comparative downstream evaluation on link prediction tasks show comparable state of the art performance to similar VGAE arrangements despite considerable simplification. We also show the simple application of our methodology to more challenging representation learning scenarios such as spatio-temporal graph representation learning.

Understanding the underlying mechanisms that enable the empirical successes of deep neural networks is essential for further improving their performance and explaining such networks. Towards this goal, a specific question is how to explain the "surprising" behavior of the same over-parametrized deep neural networks that can generalize well on real datasets and at the same time "memorize" training samples when the labels are randomized. In this paper, we demonstrate that deep ReLU networks generalize from training samples to new points via piece-wise linear interpolation. We provide a quantified analysis on the generalization ability of a deep ReLU network: Given a fixed point $\mathbf{x}$ and a fixed direction in the input space $\mathcal{S}$, there is always a segment such that any point on the segment will be classified the same as the fixed point $\mathbf{x}$. We call this segment the $generalization \ interval$. We show that the generalization intervals of a ReLU network behave similarly along pairwise directions between samples of the same label in both real and random cases on the MNIST and CIFAR-10 datasets. This result suggests that the same interpolation mechanism is used in both cases. Additionally, for datasets using real labels, such networks provide a good approximation of the underlying manifold in the data, where the changes are much smaller along tangent directions than along normal directions. On the other hand, however, for datasets with random labels, generalization intervals along mid-lines of triangles with the same label are much smaller than those on the datasets with real labels, suggesting different behaviors along other directions. Our systematic experiments demonstrate for the first time that such deep neural networks generalize through the same interpolation and explain the differences between their performance on datasets with real and random labels.

Adversarial Regression is a proposition to perform high dimensional non-linear regression with uncertainty estimation. We used Conditional Generative Adversarial Network to obtain an estimate of the full predictive distribution for a new observation. Generative Adversarial Networks (GAN) are implicit generative models which produce samples from a distribution approximating the distribution of the data. The conditional version of it (CGAN) takes the following expression: $\min\limits_G \max\limits_D V(D, G) = \mathbb{E}_{x\sim p_{r}(x)} [log(D(x, y))] + \mathbb{E}_{z\sim p_{z}(z)} [log (1-D(G(z, y)))]$. An approximate solution can be found by training simultaneously two neural networks to model D and G and feeding G with a random noise vector $z$. After training, we have that $G(z, y)\mathrel{\dot\sim} p_{data}(x, y)$. By fixing $y$, we have $G(z|y) \mathrel{\dot\sim} p{data}(x|y)$. By sampling $z$, we can therefore obtain samples following approximately $p(x|y)$, which is the predictive distribution of $x$ for a new $y$. We ran experiments to test various loss functions, data distributions, sample size, size of the noise vector, etc. Even if we observed differences, no experiment outperformed consistently the others. The quality of CGAN for regression relies on fine-tuning a range of hyperparameters. In a broader view, the results show that CGANs are very promising methods to perform uncertainty estimation for high dimensional non-linear regression.