As artificial agents proliferate, it is becoming increasingly important to ensure that their interactions with one another are well-behaved. In this paper, we formalize a common-sense notion of when algorithms are well-behaved: an algorithm is safe if it does no harm. Motivated by recent progress in deep learning, we focus on the specific case where agents update their actions according to gradient descent. The paper shows that that gradient descent converges to a Nash equilibrium in safe games. The main contribution is to define strongly-typed agents and show they are guaranteed to interact safely, thereby providing sufficient conditions to guarantee safe interactions. A series of examples show that strong-typing generalizes certain key features of convexity, is closely related to blind source separation, and introduces a new perspective on classical multilinear games based on tensor decomposition.

Layout design with complex constraints is a challenging problem to solve due to the non-uniqueness of the solution and the difficulties in incorporating the constraints into the conventional optimization-based methods. In this paper, we propose a design method based on the recently developed machine learning technique, Variational Autoencoder (VAE). We utilize the learning capability of the VAE to learn the constraints and the generative capability of the VAE to generate design candidates that automatically satisfy all the constraints. As such, no constraints need to be imposed during the design stage. In addition, we show that the VAE network is also capable of learning the underlying physics of the design problem, leading to an efficient design tool that does not need any physical simulation once the network is constructed. We demonstrated the performance of the method on two cases: inverse design of surface diffusion induced morphology change and mask design for optical microlithography.

There are many advantages to use probability method for nonlinear system identification, such as the noises and outliers in the data set do not affect the probability models significantly; the input features can be extracted in probability forms. The biggest obstacle of the probability model is the probability distributions are not easy to be obtained. In this paper, we form the nonlinear system identification into solving the conditional probability. Then we modify the restricted Boltzmann machine (RBM), such that the joint probability, input distribution, and the conditional probability can be calculated by the RBM training. Binary encoding and continue valued methods are discussed. The universal approximation analysis for the conditional probability based modelling is proposed. We use two benchmark nonlinear systems to compare our probability modelling method with the other black-box modeling methods. The results show that this novel method is much better when there are big noises and the system dynamics are complex.

Generating novel graph structures that optimize given objectives while obeying some given underlying rules is fundamental for chemistry, biology and social science research. This is especially important in the task of molecular graph generation, whose goal is to discover novel molecules with desired properties such as drug-likeness and synthetic accessibility, while obeying physical laws such as chemical valency. However, designing models to find molecules that optimize desired properties while incorporating highly complex and non-differentiable rules remains to be a challenging task. Here we propose Graph Convolutional Policy Network (GCPN), a general graph convolutional network based model for goal-directed graph generation through reinforcement learning. The model is trained to optimize domain-specific rewards and adversarial loss through policy gradient, and acts in an environment that incorporates domain-specific rules. Experimental results show that GCPN can achieve 61% improvement on chemical property optimization over state-of-the-art baselines while resembling known molecules, and achieve 184% improvement on the constrained property optimization task.

A shallow neural network is a universal function approximator, if allowed an unlimited number of neurons in its single hidden layer (Cybenko, 1989; Hornik et al., 1989). Since in theory a shallow network can do anything, what is the advantage of going deep? For one thing, deeper architectures are capable of encoding certain types of functions far more efficiently than their shallow counterparts (Montufar et al., 2014; Szymanski and McCane, 2014; Telgarsky, 2015). The efficiency of function encoding is important for two reasons: - deep learning can tackle problems that may be computationally intractable with the shallow approach; - the fewer trainable parameters in a neural network, the lower the bound on their generalisation error (Vapnik, 1998) due to decreased generic representational power (Anthony and Bartlett, 2009; Bartlett et al., 2017), While the first point is highly relevant for practical purposes, the latter is more interesting from the theoretical point of view. An approximation that can be made to the same level of accuracy with significantly fewer parameters is likely to give better generalisation.

The realization that video games are perfect testbeds for artificial intelligence methods have in recent years spread to the whole AI community, in particular since Chess and Go have been effectively conquered, and there is an almost daily flurry of new papers applying AI methods to video games. In particular, the Arcade Learning Environment (ALE), which builds on an emulator for the Atari 2600 games console and contains several dozens of games [1], have been used in numerous published papers since DeepMind's landmark paper showing that Q-learning combined with deep convolutional networks could learn to play many of the ALE games at superhuman level [2]. As an AI benchmark, ALE is limited in the sense that there is only a finite set of games. This is a limitation it has in common with any framework based on existing published games. However, for being able to test the general video game playing ability of an agent, it is necessary to test on games on which the agent was not optimized.

Many real-world sequential decision making problems are partially observable by nature, and the environment model is typically unknown. Consequently, there is great need for reinforcement learning methods that can tackle such problems given only a stream of incomplete and noisy observations. In this paper, we propose deep variational reinforcement learning (DVRL), which introduces an inductive bias that allows an agent to learn a generative model of the environment and perform inference in that model to effectively aggregate the available information. We develop an n-step approximation to the evidence lower bound (ELBO), allowing the model to be trained jointly with the policy. This ensures that the latent state representation is suitable for the control task. In experiments on Mountain Hike and flickering Atari we show that our method outperforms previous approaches relying on recurrent neural networks to encode the past.

This paper introduces the variational implicit processes (VIPs), a Bayesian nonparametric method based on a class of highly flexible priors over functions. Similar to Gaussian processes (GPs), in implicit processes (IPs), an implicit multivariate prior (data simulators, Bayesian neural networks, etc.) is placed over any finite collections of random variables. A novel and efficient variational inference algorithm for IPs is derived using wake-sleep updates, which gives analytic solutions and allows scalable hyper-parameter learning with stochastic optimization. Experiments on real-world regression datasets demonstrate that VIPs return better uncertainty estimates and superior performance over existing inference methods for GPs and Bayesian neural networks. With a Bayesian LSTM as the implicit prior, the proposed approach achieves state-of-the-art results on predicting power conversion efficiency of molecules based on raw chemical formulas.

We propose a single neural probabilistic model based on variational autoencoder that can be conditioned on an arbitrary subset of observed features and then sample the remaining features in "one shot". The features may be both real-valued and categorical. Training of the model is performed by stochastic variational Bayes. The experimental evaluation on synthetic data, as well as feature imputation and image inpainting problems, shows the effectiveness of the proposed approach and diversity of the generated samples.

Deep learning on graph structures has shown exciting results in various applications. However, few attentions have been paid to the robustness of such models, in contrast to numerous research work for image or text adversarial attack and defense. In this paper, we focus on the adversarial attacks that fool the model by modifying the combinatorial structure of data. We first propose a reinforcement learning based attack method that learns the generalizable attack policy, while only requiring prediction labels from the target classifier. Also, variants of genetic algorithms and gradient methods are presented in the scenario where prediction confidence or gradients are available. We use both synthetic and real-world data to show that, a family of Graph Neural Network models are vulnerable to these attacks, in both graph-level and node-level classification tasks. We also show such attacks can be used to diagnose the learned classifiers.