Scalable Data Augmentation (SDA) provides a framework for training deep neural networks (DNNs). Our methodology exploits auxiliary hidden units which are designed to avoid backtracking and traverse local modes in an efficient way. This allows us to exploit recent advantages in high performance computing such as scalable linear algebra (CUDA, XLA). We show how to implement standard activation and objective functions, including ReLU (Polson and Ročková, 2018), logit (Zhou et al., 2012) and SVM (Mallick et al., 2005) are all available as data augmentation schemes. Data augmentation strategies are commonplace in statistical applications such as EM, ECM and MM algorithms, as they accelerate convergence and can use Nesterov acceleration (Nesterov, 1983).

Step sizes in neural network training are largely determined using predetermined rules such as fixed learning rates and learning rate schedules, which require user input to determine their functional form and associated hyperparameters. Global optimization strategies to resolve these hyperparameters are computationally expensive. Line searches are capable of adaptively resolving learning rate schedules. However, due to discontinuities induced by mini-batch sampling, they have largely fallen out of favor. Notwithstanding, probabilistic line searches have recently demonstrated viability in resolving learning rates for stochastic loss functions. This method creates surrogates with confidence intervals, where restrictions are placed on the rate at which the search domain can grow along a search direction. This paper introduces an alternative paradigm, Gradient-Only Line Searches that are inexact (GOLS-I), as an alternative strategy to automatically resolve learning rates in stochastic cost functions over a range of 15 orders of magnitude without the use of surrogates. We show that GOLS-I is a competitive strategy to reliably resolve step sizes, adding high value in terms of performance, while being easy to implement. Considering mini-batch sampling, we open the discussion on how to split the effort to resolve quality search directions from quality step size estimates along a search direction.

For autonomous agents to successfully operate in real world, the ability to anticipate future motions of surrounding entities in the scene can greatly enhance their safety levels since potentially dangerous situations could be avoided in advance. While impressive results have been shown on predicting each agent's behavior independently, we argue that it is not valid to consider road entities individually since transitions of vehicle states are highly coupled. Moreover, as the predicted horizon becomes longer, modeling prediction uncertainties and multi-modal distributions over future sequences will turn into a more challenging task. In this paper, we address this challenge by presenting a multi-modal probabilistic prediction approach. The proposed method is based on a generative model and is capable of jointly predicting sequential motions of each pair of interacting agents. Most importantly, our model is interpretable, which can explain the underneath logic as well as obtain more reliability to use in real applications. A complicate real-world roundabout scenario is utilized to implement and examine the proposed method.

In stochastic differential games, a Nash equilibrium refers to strategies by which no player has an incentive to deviate. Finding a Nash equilibrium is one of the core problems in noncooperative game theory, however, due to the notorious intractability of N-player game, the computation of the Nash equilibrium has been shown extremely time-consuming and memory demanding, especially for large N [16]. On the other hand, a rich literature on game theory has been developed to study consequences of strategies on interactions between a large group of rational "agents", e.g., system risk caused by inter-bank borrowing and lending, price impacts imposed by agents' optimal liquidation, and market price from monopolistic competition. This makes it crucial to develop efficient theory and fast algorithms for computing the Nash equilibrium of N-player stochastic differential games. Deep neural networks with many layers have been recently shown to do a great job in artificial intelligence (e.g., [2, 39]). The idea behind is to use compositions of simple functions to approximate complicated ones, and there are approximation theorems showing that a wide class of functions on compact subsets can be approximated by a single hidden layer neural network (e.g., [53]). This brings a possibility of solving a high-dimensional system using deep neural networks, and in fact, these techniques have been successfully applied to solve stochastic control problems [20, 29, 1]. In this paper, we propose to build deep neural networks by using strategies of fictitious play, and develop deep learning algorithms for computing the Nash equilibrium of asymmetric N-player non-zerosum stochastic differential games.

The recommender system is an important form of intelligent application, which assists users to alleviate from information redundancy. Among the metrics used to evaluate a recommender system, the metric of conversion has become more and more important. The majority of existing recommender systems perform poorly on the metric of conversion due to its extremely sparse feedback signal. To tackle this challenge, we propose a deep hierarchical reinforcement learning based recommendation framework, which consists of two components, i.e., high-level agent and low-level agent. The high-level agent catches long-term sparse conversion signals, and automatically sets abstract goals for low-level agent, while the low-level agent follows the abstract goals and interacts with real-time environment. To solve the inherent problem in hierarchical reinforcement learning, we propose a novel deep hierarchical reinforcement learning algorithm via multi-goals abstraction (HRL-MG). Our proposed algorithm contains three characteristics: 1) the high-level agent generates multiple goals to guide the low-level agent in different stages, which reduces the difficulty of approaching high-level goals; 2) different goals share the same state encoder parameters, which increases the update frequency of the high-level agent and thus accelerates the convergence of our proposed algorithm; 3) an appreciate benefit assignment function is designed to allocate rewards in each goal so as to coordinate different goals in a consistent direction. We evaluate our proposed algorithm based on a real-world e-commerce dataset and validate its effectiveness.

One problem in the application of reinforcement learning to real-world problems is the curse of dimensionality on the action space. Macro actions, a sequence of primitive actions, have been studied to diminish the dimensionality of the action space with regard to the time axis. However, previous studies relied on humans defining macro actions or assumed macro actions as repetitions of the same primitive actions. We present Factorized Macro Action Reinforcement Learning (FaMARL) which autonomously learns disentangled factor representation of a sequence of actions to generate macro actions that can be directly applied to general reinforcement learning algorithms. FaMARL exhibits higher scores than other reinforcement learning algorithms on environments that require an extensive amount of search.

DeepMind's Research Platform Team has open-sourced TF-Replicator, a framework that enables researchers without previous experience with the distributed system to deploy their TensorFlow models on GPUs and Cloud TPUs. The move aims to strengthen AI research and development. Synced invited Yuan Tang, a senior software engineer at Ant Financial, to share his thoughts on TF-Replicator. How would you describe TF-Replicator? TF-Replicator is a framework to simplify the writing of distributed TensorFlow code for training machine learning models, so that they can be effortlessly deployed to different cluster architectures.

In classification problems (i.e. when the labels Y are discrete), the scaling of the test error with respect to n is determined by characterizations of the VC-dimension [2]/Rademacher complexity [3] of the function class, which in the worst case increases with its number of parameters. In regression (i.e. when the labels Y are continuous), the mean-squared error of the ordinary least-squares estimate is characterized by the condition number of the regression matrix, which is reasonable for appropriate ratios of d/n but tends to increase astronomically as d approaches n. The qualitative fear is the same: if the function class is too complex, it starts to overfit noise and can generalize poorly to unseen test data. But there is a gap between "can" and "will" -- and indeed this conventional wisdom has been challenged by the recent advent of deeper and deeper neural networks. In particular, a thought-provoking paper [4] noted that several deep neural networks generalize well despite achieving zero or close to zero training error, and being so expressive that they even have the ability to fit pure noise. As they put it, "understanding deep learning requires rethinking generalization". How can we reconcile the fact that good interpolative solutions exist with the classical bias-variance tradeoff? These phenomena are being actively investigated in a statistical sense [5,6] and a computational sense [7-9] in classification problems and/or noiseless models.

This paper proposes using a linear function approximator, rather than a deep neural network (DNN), to bias a Monte Carlo tree search (MCTS) player for general games. This is unlikely to match the potential raw playing strength of DNNs, but has advantages in terms of generality, interpretability and resources (time and hardware) required for training. Features describing local patterns are used as inputs. The features are formulated in such a way that they are easily interpretable and applicable to a wide range of general games, and might encode simple local strategies. We gradually create new features during the same self-play training process used to learn feature weights. We evaluate the playing strength of an MCTS player biased by learnt features against a standard upper confidence bounds for trees (UCT) player in multiple different board games, and demonstrate significantly improved playing strength in the majority of them after a small number of self-play training games.

An environment representation (ER) is a substantial part of every autonomous system. It introduces a common interface between perception and other system components, such as decision making, and allows downstream algorithms to deal with abstracted data without knowledge of the used sensor. In this work, we propose and evaluate a novel architecture that generates an egocentric, grid-based, predictive, and semantically-interpretable ER. In particular, we provide a proof of concept for the spatio-temporal fusion of multiple camera sequences and short-term prediction in such an ER. Our design utilizes a strong semantic segmentation network together with depth and egomotion estimates to first extract semantic information from multiple camera streams and then transform these separately into egocentric temporally-aligned bird's-eye view grids. A deep encoder-decoder network is trained to fuse a stack of these grids into a unified semantic grid representation and to predict the dynamics of its surrounding. We evaluate this representation on real-world sequences of the Cityscapes dataset and show that our architecture can make accurate predictions in complex sensor fusion scenarios and significantly outperforms a model-driven baseline in a category-based evaluation.