Publications

David Silver

Nicolas Heess

Value estimation is a critical component of the reinforcement learning (RL) paradigm. The question of how to effectively learn predictors fo… (see more)r value from data is one of the major problems studied by the RL community, and different approaches exploit structure in the problem domain in different ways. Model learning can make use of the rich transition structure present in sequences of observations, but this approach is usually not sensitive to the reward function. In contrast, model-free methods directly leverage the quantity of interest from the future but have to compose with a potentially weak scalar signal (an estimate of the return). In this paper we develop an approach for representation learning in RL that sits in between these two extremes: we propose to learn what to model in a way that can directly help value prediction. To this end we determine which features of the future trajectory provide useful information to predict the associated return. This provides us with tractable prediction targets that are directly relevant for a task, and can thus accelerate learning of the value function. The idea can be understood as reasoning, in hindsight, about which aspects of the future observations could help past value prediction. We show how this can help dramatically even in simple policy evaluation settings. We then test our approach at scale in challenging domains, including on 57 Atari 2600 games.

2019-12-31

Advances in Neural Information Processing Systems 33 (NeurIPS 2020) (published)

openreview.net

Learning from Learning Machines: Optimisation, Rules, and Social Norms

Travis LaCroix

Yoshua Bengio

There is an analogy between machine learning systems and economic entities in that they are both adaptive, and their behaviour is specified … (see more)in a more-or-less explicit way. It appears that the area of AI that is most analogous to the behaviour of economic entities is that of morally good decision-making, but it is an open question as to how precisely moral behaviour can be achieved in an AI system. This paper explores the analogy between these two complex systems, and we suggest that a clearer understanding of this apparent analogy may help us forward in both the socio-economic domain and the AI domain: known results in economics may help inform feasible solutions in AI safety, but also known results in AI may inform economic policy. If this claim is correct, then the recent successes of deep learning for AI suggest that more implicit specifications work better than explicit ones for solving such problems.

2019-12-28

ArXiv (preprint)

Assessment of Extubation Readiness Using Spontaneous Breathing Trials in Extremely Preterm Neonates

Wissam Shalish

Lara Kanbar

Lajos Kovacs

Sanjay Chawla

Martin Keszler

Smita Rao

Samantha Latremouille

Karen Brown

Robert E. Kearney

Guilherme M. Sant’Anna

Importance Spontaneous breathing trials (SBTs) are used to determine extubation readiness in extremely preterm neonates (gestational age ≤… (see more)28 weeks), but these trials rely on empirical combinations of clinical events during endotracheal continuous positive airway pressure (ET-CPAP). Objectives To describe clinical events during ET-CPAP and to assess accuracy of comprehensive clinical event combinations in predicting successful extubation compared with clinical judgment alone. Design, Setting, and Participants This multicenter diagnostic study used data from 259 neonates seen at 5 neonatal intensive care units from the prospective Automated Prediction of Extubation Readiness (APEX) study from September 1, 2013, through August 31, 2018. Neonates with birth weight less than 1250 g who required mechanical ventilation were eligible. Neonates deemed to be ready for extubation and who underwent ET-CPAP before extubation were included. Interventions In the APEX study, cardiorespiratory signals were recorded during 5-minute ET-CPAP, and signs of clinical instability were monitored. Main Outcomes and Measures Four clinical events were documented during ET-CPAP: apnea requiring stimulation, presence and cumulative durations of bradycardia and desaturation, and increased supplemental oxygen. Clinical event occurrence was assessed and compared between extubation pass and fail (defined as reintubation within 7 days). An automated algorithm was developed to generate SBT definitions using all clinical event combinations and to compute diagnostic accuracies of an SBT in predicting extubation success. Results Of 259 neonates (139 [54%] male) with a median gestational age of 26.1 weeks (interquartile range [IQR], 24.9-27.4 weeks) and median birth weight of 830 g (IQR, 690-1019 g), 147 (57%) had at least 1 clinical event during ET-CPAP. Apneas occurred in 10% (26 of 259) of neonates, bradycardias in 19% (48), desaturations in 53% (138), and increased oxygen needs in 41% (107). Neonates with successful extubation (71% [184 of 259]) had significantly fewer clinical events (51% [93 of 184] vs 72% [54 of 75], P = .002), shorter cumulative bradycardia duration (median, 0 seconds [IQR, 0 seconds] vs 0 seconds [IQR, 0-9 seconds], P .001), shorter cumulative desaturation duration (median, 0 seconds [IQR, 0-59 seconds] vs 25 seconds [IQR, 0-90 seconds], P = .003),

2019-12-19

JAMA Pediatrics (published)

CLOSURE: Assessing Systematic Generalization of CLEVR Models

Dzmitry Bahdanau

Harm de Vries

Timothy J. O'Donnell

Shikhar Murty

Philippe Beaudoin

Yoshua Bengio

Aaron Courville

The CLEVR dataset of natural-looking questions about 3D-rendered scenes has recently received much attention from the research community. A … (see more)number of models have been proposed for this task, many of which achieved very high accuracies of around 97-99%. In this work, we study how systematic the generalization of such models is, that is to which extent they are capable of handling novel combinations of known linguistic constructs. To this end, we test models' understanding of referring expressions based on matching object properties (such as e.g. "the object that is the same size as the red ball") in novel contexts. Our experiments on the thereby constructed CLOSURE benchmark show that state-of-the-art models often do not exhibit systematicity after being trained on CLEVR. Surprisingly, we find that an explicitly compositional Neural Module Network model also generalizes badly on CLOSURE, even when it has access to the ground-truth programs at test time. We improve the NMN's systematic generalization by developing a novel Vector-NMN module architecture with vector-valued inputs and outputs. Lastly, we investigate the extent to which few-shot transfer learning can help models that are pretrained on CLEVR to adapt to CLOSURE. Our few-shot learning experiments contrast the adaptation behavior of the models with intermediate discrete programs with that of the end-to-end continuous models.

2019-12-11

ArXiv (preprint)

Shaping representations through communication: community size effect in artificial learning systems

Olivier Tieleman

Angeliki Lazaridou

Shibl Mourad

Charles Blundell

Motivated by theories of language and communication that explain why communities with large numbers of speakers have, on average, simpler la… (see more)nguages with more regularity, we cast the representation learning problem in terms of learning to communicate. Our starting point sees the traditional autoencoder setup as a single encoder with a fixed decoder partner that must learn to communicate. Generalizing from there, we introduce community-based autoencoders in which multiple encoders and decoders collectively learn representations by being randomly paired up on successive training iterations. We find that increasing community sizes reduce idiosyncrasies in the learned codes, resulting in representations that better encode concept categories and correlate with human feature norms.

2019-12-11

ArXiv (preprint)

Approximate information state for partially observed systems

Jayakumar Subramanian

Aditya Mahajan

The standard approach for modeling partially observed systems is to model them as partially observable Markov decision processes (POMDPs) an… (see more)d obtain a dynamic program in terms of a belief state. The belief state formulation works well for planning but is not ideal for online reinforcement learning because the belief state depends on the model and, as such, is not observable when the model is unknown.In this paper, we present an alternative notion of an information state for obtaining a dynamic program in partially observed models. In particular, an information state is a sufficient statistic for the current reward which evolves in a controlled Markov manner. We show that such an information state leads to a dynamic programming decomposition. Then we present a notion of an approximate information state and present an approximate dynamic program based on the approximate information state. Approximate information state is defined in terms of properties that can be estimated using sampled trajectories. Therefore, they provide a constructive method for reinforcement learning in partially observed systems. We present one such construction and show that it performs better than the state of the art for three benchmark models.

2019-12-10

2019 IEEE 58th Conference on Decision and Control (CDC) (published)

Entropy Regularization with Discounted Future State Distribution in Policy Gradient Methods

The policy gradient theorem is defined based on an objective with respect to the initial distribution over states. In the discounted case, t… (see more)his results in policies that are optimal for one distribution over initial states, but may not be uniformly optimal for others, no matter where the agent starts from. Furthermore, to obtain unbiased gradient estimates, the starting point of the policy gradient estimator requires sampling states from a normalized discounted weighting of states. However, the difficulty of estimating the normalized discounted weighting of states, or the stationary state distribution, is quite well-known. Additionally, the large sample complexity of policy gradient methods is often attributed to insufficient exploration, and to remedy this, it is often assumed that the restart distribution provides sufficient exploration in these algorithms. In this work, we propose exploration in policy gradient methods based on maximizing entropy of the discounted future state distribution. The key contribution of our work includes providing a practically feasible algorithm to estimate the normalized discounted weighting of states, i.e, the \textit{discounted future state distribution}. We propose that exploration can be achieved by entropy regularization with the discounted state distribution in policy gradients, where a metric for maximal coverage of the state space can be based on the entropy of the induced state distribution. The proposed approach can be considered as a three time-scale algorithm and under some mild technical conditions, we prove its convergence to a locally optimal policy. Experimentally, we demonstrate usefulness of regularization with the discounted future state distribution in terms of increased state space coverage and faster learning on a range of complex tasks.

2019-12-10

ArXiv (preprint)

Marginalized State Distribution Entropy Regularization in Policy Optimization

Riashat Islam

Zafarali Ahmed

2019-12-10

ArXiv (preprint)

Networked control of coupled subsystems: Spectral decomposition and low-dimensional solutions

Shuang Gao

Aditya Mahajan

In this paper, we investigate optimal networked control of coupled subsystems where the dynamics and the cost couplings depend on an underly… (see more)ing weighted graph. We use the spectral decomposition of the graph adjacency matrix to decompose the overall system into (L+1) systems with decoupled dynamics and cost, where L is the rank of the adjacency matrix. Consequently, the optimal control input at each subsystem can be computed by solving (L+1) decoupled Riccati equations. A salient feature of the result is that the solution complexity depends on the rank of the adjacency matrix rather than the size of the network (i.e., the number of nodes). Therefore, the proposed solution framework provides a scalable method for synthesizing and implementing optimal control laws for large-scale systems.

2019-12-10

2019 IEEE 58th Conference on Decision and Control (CDC) (published)

Restless bandits with controlled restarts: Indexability and computation of Whittle index

Nima Akbarzadeh

Aditya Mahajan

Motivated by applications in machine repair, queueing, surveillance, and clinic care, we consider a scheduling problem where a decision make… (see more)r can reset m out of n Markov processes at each time. Processes that are reset, restart according to a known probability distribution and processes that are not reset, evolve in a Markovian manner. Due to the high complexity of finding an optimal policy, such scheduling problems are often modeled as restless bandits. We show that the model satisfies a technical condition known as indexability. For indexable restless bandits, the Whittle index policy, which computes a function known as Whittle index for each process and resets the m processes with the lowest index, is known to be a good heuristic. The Whittle index is computed by solving an auxiliary Markov decision problem for each arm. When the optimal policy for this auxiliary problem is threshold based, we use ideas from renewal theory to derive closed form expression for the Whittle index. We present detailed numerical experiments which suggest that Whittle index policy performs close to the optimal policy and performs significantly better than myopic policy, which is a commonly used heuristic.

2019-12-10

2019 IEEE 58th Conference on Decision and Control (CDC) (published)