Kai Arulkumaran

Arthur Juliani, Kai Arulkumaran, Shuntaro Sasai et al.

In popular media, there is often a connection drawn between the advent of awareness in artificial agents and those same agents simultaneously achieving human or superhuman level intelligence. In this work, we explore the validity and potential application of this seemingly intuitive link between consciousness and intelligence. We do so by examining the cognitive abilities associated with three contemporary theories of conscious function: Global Workspace Theory (GWT), Information Generation Theory (IGT), and Attention Schema Theory (AST). We find that all three theories specifically relate conscious function to some aspect of domain-general intelligence in humans. With this insight, we turn to the field of Artificial Intelligence (AI) and find that, while still far from demonstrating general intelligence, many state-of-the-art deep learning methods have begun to incorporate key aspects of each of the three functional theories. Having identified this trend, we use the motivating example of mental time travel in humans to propose ways in which insights from each of the three theories may be combined into a single unified and implementable model. Given that it is made possible by cognitive abilities underlying each of the three functional theories, artificial agents capable of mental time travel would not only possess greater general intelligence than current approaches, but also be more consistent with our current understanding of the functional role of consciousness in humans, thus making it a promising near-term goal for AI research.

Shivakanth Sujit, Luca Nunziante, Dan Ogawa Lillrank et al.

In this work we extend the low-cost GELLO teleoperation system, initially designed for joint position control, with additional force information. Our first extension is to implement force feedback, allowing users to feel resistance when interacting with the environment. Our second extension is to add force information into the data collection process and training of imitation learning models. We validate our additions by implementing these on a GELLO system with a Franka Panda arm as the follower robot, performing a user study, and comparing the performance of policies trained with and without force information on a range of simulated and real dexterous manipulation tasks. Qualitatively, users with robotics experience preferred our controller, and the addition of force inputs improved task success on the majority of tasks.

Masakazu Inoue, Motoshige Sato, Kenichi Tomeoka et al.

Silent speech decoding, which performs unvocalized human speech recognition from electroencephalography/electromyography (EEG/EMG), increases accessibility for speech-impaired humans. However, data collection is difficult and performed using varying experimental setups, making it nontrivial to collect a large, homogeneous dataset. In this study we introduce neural networks that can handle EEG/EMG with heterogeneous electrode placements and show strong performance in silent speech decoding via multi-task training on large-scale EEG/EMG datasets. We achieve improved word classification accuracy in both healthy participants (95.3%), and a speech-impaired patient (54.5%), substantially outperforming models trained on single-subject data (70.1% and 13.2%). Moreover, our models also show gains in cross-language calibration performance. This increase in accuracy suggests the feasibility of developing practical silent speech decoding systems, particularly for speech-impaired patients.

Kai Arulkumaran, Dylan R. Ashley, Jürgen Schmidhuber et al.

Upside down reinforcement learning (UDRL) flips the conventional use of the return in the objective function in RL upside down, by taking returns as input and predicting actions. UDRL is based purely on supervised learning, and bypasses some prominent issues in RL: bootstrapping, off-policy corrections, and discount factors. While previous work with UDRL demonstrated it in a traditional online RL setting, here we show that this single algorithm can also work in the imitation learning and offline RL settings, be extended to the goal-conditioned RL setting, and even the meta-RL setting. With a general agent architecture, a single UDRL agent can learn across all paradigms.

Dylan R. Ashley, Kai Arulkumaran, Jürgen Schmidhuber et al.

Lately, there has been a resurgence of interest in using supervised learning to solve reinforcement learning problems. Recent work in this area has largely focused on learning command-conditioned policies. We investigate the potential of one such method -- upside-down reinforcement learning -- to work with commands that specify a desired relationship between some scalar value and the observed return. We show that upside-down reinforcement learning can learn to carry out such commands online in a tabular bandit setting and in CartPole with non-linear function approximation. By doing so, we demonstrate the power of this family of methods and open the way for their practical use under more complicated command structures.

Tianhong Dai, Hengyan Liu, Kai Arulkumaran et al.

Hindsight experience replay (HER) is a goal relabelling technique typically used with off-policy deep reinforcement learning algorithms to solve goal-oriented tasks; it is well suited to robotic manipulation tasks that deliver only sparse rewards. In HER, both trajectories and transitions are sampled uniformly for training. However, not all of the agent's experiences contribute equally to training, and so naive uniform sampling may lead to inefficient learning. In this paper, we propose diversity-based trajectory and goal selection with HER (DTGSH). Firstly, trajectories are sampled according to the diversity of the goal states as modelled by determinantal point processes (DPPs). Secondly, transitions with diverse goal states are selected from the trajectories by using k-DPPs. We evaluate DTGSH on five challenging robotic manipulation tasks in simulated robot environments, where we show that our method can learn more quickly and reach higher performance than other state-of-the-art approaches on all tasks.

Kai Arulkumaran, Dan Ogawa Lillrank

The introduction of the generative adversarial imitation learning (GAIL) algorithm has spurred the development of scalable imitation learning approaches using deep neural networks. Many of the algorithms that followed used a similar procedure, combining on-policy actor-critic algorithms with inverse reinforcement learning. More recently there have been an even larger breadth of approaches, most of which use off-policy algorithms. However, with the breadth of algorithms, everything from datasets to base reinforcement learning algorithms to evaluation settings can vary, making it difficult to fairly compare them. In this work we re-implement 6 different IL algorithms, updating 3 of them to be off-policy, base them on a common off-policy algorithm (SAC), and evaluate them on a widely-used expert trajectory dataset (D4RL) for the most common benchmark (MuJoCo). After giving all algorithms the same hyperparameter optimisation budget, we compare their results for a range of expert trajectories. In summary, GAIL, with all of its improvements, consistently performs well across a range of sample sizes, AdRIL is a simple contender that performs well with one important hyperparameter to tune, and behavioural cloning remains a strong baseline when data is more plentiful.

Pierre-Alexandre Kamienny, Kai Arulkumaran, Feryal Behbahani et al.

Using privileged information during training can improve the sample efficiency and performance of machine learning systems. This paradigm has been applied to reinforcement learning (RL), primarily in the form of distillation or auxiliary tasks, and less commonly in the form of augmenting the inputs of agents. In this work, we investigate Privileged Information Dropout (\pid) for achieving the latter which can be applied equally to value-based and policy-based RL algorithms. Within a simple partially-observed environment, we demonstrate that \pid outperforms alternatives for leveraging privileged information, including distillation and auxiliary tasks, and can successfully utilise different types of privileged information. Finally, we analyse its effect on the learned representations.

Tianhong Dai, Kai Arulkumaran, Tamara Gerbert et al.

Deep reinforcement learning has the potential to train robots to perform complex tasks in the real world without requiring accurate models of the robot or its environment. A practical approach is to train agents in simulation, and then transfer them to the real world. One popular method for achieving transferability is to use domain randomisation, which involves randomly perturbing various aspects of a simulated environment in order to make trained agents robust to the reality gap. However, less work has gone into understanding such agents - which are deployed in the real world - beyond task performance. In this work we examine such agents, through qualitative and quantitative comparisons between agents trained with and without visual domain randomisation. We train agents for Fetch and Jaco robots on a visuomotor control task and evaluate how well they generalise using different testing conditions. Finally, we investigate the internals of the trained agents by using a suite of interpretability techniques. Our results show that the primary outcome of domain randomisation is more robust, entangled representations, accompanied with larger weights with greater spatial structure; moreover, the types of changes are heavily influenced by the task setup and presence of additional proprioceptive inputs. Additionally, we demonstrate that our domain randomised agents require higher sample complexity, can overfit and more heavily rely on recurrent processing. Furthermore, even with an improved saliency method introduced in this work, we show that qualitative studies may not always correspond with quantitative measures, necessitating the combination of inspection tools in order to provide sufficient insights into the behaviour of trained agents.

Marta Sarrico, Kai Arulkumaran, Andrea Agostinelli et al.

Deep networks have enabled reinforcement learning to scale to more complex and challenging domains, but these methods typically require large quantities of training data. An alternative is to use sample-efficient episodic control methods: neuro-inspired algorithms which use non-/semi-parametric models that predict values based on storing and retrieving previously experienced transitions. One way to further improve the sample efficiency of these approaches is to use more principled exploration strategies. In this work, we therefore propose maximum entropy mellowmax episodic control (MEMEC), which samples actions according to a Boltzmann policy with a state-dependent temperature. We demonstrate that MEMEC outperforms other uncertainty- and softmax-based exploration methods on classic reinforcement learning environments and Atari games, achieving both more rapid learning and higher final rewards.

Andrea Agostinelli, Kai Arulkumaran, Marta Sarrico et al.

Recently, neuro-inspired episodic control (EC) methods have been developed to overcome the data-inefficiency of standard deep reinforcement learning approaches. Using non-/semi-parametric models to estimate the value function, they learn rapidly, retrieving cached values from similar past states. In realistic scenarios, with limited resources and noisy data, maintaining meaningful representations in memory is essential to speed up the learning and avoid catastrophic forgetting. Unfortunately, EC methods have a large space and time complexity. We investigate different solutions to these problems based on prioritising and ranking stored states, as well as online clustering techniques. We also propose a new dynamic online k-means algorithm that is both computationally-efficient and yields significantly better performance at smaller memory sizes; we validate this approach on classic reinforcement learning environments and Atari games.

Kai Arulkumaran, Antoine Cully, Julian Togelius

In January 2019, DeepMind revealed AlphaStar to the world-the first artificial intelligence (AI) system to beat a professional player at the game of StarCraft II-representing a milestone in the progress of AI. AlphaStar draws on many areas of AI research, including deep learning, reinforcement learning, game theory, and evolutionary computation (EC). In this paper we analyze AlphaStar primarily through the lens of EC, presenting a new look at the system and relating it to many concepts in the field. We highlight some of its most interesting aspects-the use of Lamarckian evolution, competitive co-evolution, and quality diversity. In doing so, we hope to provide a bridge between the wider EC community and one of the most significant AI systems developed in recent times.

Ryutaro Tanno, Kai Arulkumaran, Daniel C. Alexander et al.

Deep neural networks and decision trees operate on largely separate paradigms; typically, the former performs representation learning with pre-specified architectures, while the latter is characterised by learning hierarchies over pre-specified features with data-driven architectures. We unite the two via adaptive neural trees (ANTs) that incorporates representation learning into edges, routing functions and leaf nodes of a decision tree, along with a backpropagation-based training algorithm that adaptively grows the architecture from primitive modules (e.g., convolutional layers). We demonstrate that, whilst achieving competitive performance on classification and regression datasets, ANTs benefit from (i) lightweight inference via conditional computation, (ii) hierarchical separation of features useful to the task e.g. learning meaningful class associations, such as separating natural vs. man-made objects, and (iii) a mechanism to adapt the architecture to the size and complexity of the training dataset.

Sebastian Tschiatschek, Kai Arulkumaran, Jan Stühmer et al.

Partially observable Markov decision processes (POMDPs) are a powerful abstraction for tasks that require decision making under uncertainty, and capture a wide range of real world tasks. Today, effective planning approaches exist that generate effective strategies given black-box models of a POMDP task. Yet, an open question is how to acquire accurate models for complex domains. In this paper we propose DELIP, an approach to model learning for POMDPs that utilizes amortized structured variational inference. We empirically show that our model leads to effective control strategies when coupled with state-of-the-art planners. Intuitively, model-based approaches should be particularly beneficial in environments with changing reward structures, or where rewards are initially unknown. Our experiments confirm that DELIP is particularly effective in this setting.

Antonia Creswell, Tom White, Vincent Dumoulin et al.

Generative adversarial networks (GANs) provide a way to learn deep representations without extensively annotated training data. They achieve this through deriving backpropagation signals through a competitive process involving a pair of networks. The representations that can be learned by GANs may be used in a variety of applications, including image synthesis, semantic image editing, style transfer, image super-resolution and classification. The aim of this review paper is to provide an overview of GANs for the signal processing community, drawing on familiar analogies and concepts where possible. In addition to identifying different methods for training and constructing GANs, we also point to remaining challenges in their theory and application.

Antonia Creswell, Kai Arulkumaran, Anil A. Bharath

Denoising autoencoders (DAEs) are powerful deep learning models used for feature extraction, data generation and network pre-training. DAEs consist of an encoder and decoder which may be trained simultaneously to minimise a loss (function) between an input and the reconstruction of a corrupted version of the input. There are two common loss functions used for training autoencoders, these include the mean-squared error (MSE) and the binary cross-entropy (BCE). When training autoencoders on image data a natural choice of loss function is BCE, since pixel values may be normalised to take values in [0,1] and the decoder model may be designed to generate samples that take values in (0,1). We show theoretically that DAEs trained to minimise BCE may be used to take gradient steps in the data space towards regions of high probability under the data-generating distribution. Previously this had only been shown for DAEs trained using MSE. As a consequence of the theory, iterative application of a trained DAE moves a data sample from regions of low probability to regions of higher probability under the data-generating distribution. Firstly, we validate the theory by showing that novel data samples, consistent with the training data, may be synthesised when the initial data samples are random noise. Secondly, we motivate the theory by showing that initial data samples synthesised via other methods may be improved via iterative application of a trained DAE to those initial samples.

Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage et al.

Deep reinforcement learning is poised to revolutionise the field of AI and represents a step towards building autonomous systems with a higher level understanding of the visual world. Currently, deep learning is enabling reinforcement learning to scale to problems that were previously intractable, such as learning to play video games directly from pixels. Deep reinforcement learning algorithms are also applied to robotics, allowing control policies for robots to be learned directly from camera inputs in the real world. In this survey, we begin with an introduction to the general field of reinforcement learning, then progress to the main streams of value-based and policy-based methods. Our survey will cover central algorithms in deep reinforcement learning, including the deep $Q$-network, trust region policy optimisation, and asynchronous advantage actor-critic. In parallel, we highlight the unique advantages of deep neural networks, focusing on visual understanding via reinforcement learning. To conclude, we describe several current areas of research within the field.

Nat Dilokthanakul, Pedro A. M. Mediano, Marta Garnelo et al.

We study a variant of the variational autoencoder model (VAE) with a Gaussian mixture as a prior distribution, with the goal of performing unsupervised clustering through deep generative models. We observe that the known problem of over-regularisation that has been shown to arise in regular VAEs also manifests itself in our model and leads to cluster degeneracy. We show that a heuristic called minimum information constraint that has been shown to mitigate this effect in VAEs can also be applied to improve unsupervised clustering performance with our model. Furthermore we analyse the effect of this heuristic and provide an intuition of the various processes with the help of visualizations. Finally, we demonstrate the performance of our model on synthetic data, MNIST and SVHN, showing that the obtained clusters are distinct, interpretable and result in achieving competitive performance on unsupervised clustering to the state-of-the-art results.

Antonia Creswell, Kai Arulkumaran, Anil Anthony Bharath

We focus on generative autoencoders, such as variational or adversarial autoencoders, which jointly learn a generative model alongside an inference model. Generative autoencoders are those which are trained to softly enforce a prior on the latent distribution learned by the inference model. We call the distribution to which the inference model maps observed samples, the learned latent distribution, which may not be consistent with the prior. We formulate a Markov chain Monte Carlo (MCMC) sampling process, equivalent to iteratively decoding and encoding, which allows us to sample from the learned latent distribution. Since, the generative model learns to map from the learned latent distribution, rather than the prior, we may use MCMC to improve the quality of samples drawn from the generative model, especially when the learned latent distribution is far from the prior. Using MCMC sampling, we are able to reveal previously unseen differences between generative autoencoders trained either with or without a denoising criterion.

Marta Garnelo, Kai Arulkumaran, Murray Shanahan

Deep reinforcement learning (DRL) brings the power of deep neural networks to bear on the generic task of trial-and-error learning, and its effectiveness has been convincingly demonstrated on tasks such as Atari video games and the game of Go. However, contemporary DRL systems inherit a number of shortcomings from the current generation of deep learning techniques. For example, they require very large datasets to work effectively, entailing that they are slow to learn even when such datasets are available. Moreover, they lack the ability to reason on an abstract level, which makes it difficult to implement high-level cognitive functions such as transfer learning, analogical reasoning, and hypothesis-based reasoning. Finally, their operation is largely opaque to humans, rendering them unsuitable for domains in which verifiability is important. In this paper, we propose an end-to-end reinforcement learning architecture comprising a neural back end and a symbolic front end with the potential to overcome each of these shortcomings. As proof-of-concept, we present a preliminary implementation of the architecture and apply it to several variants of a simple video game. We show that the resulting system -- though just a prototype -- learns effectively, and, by acquiring a set of symbolic rules that are easily comprehensible to humans, dramatically outperforms a conventional, fully neural DRL system on a stochastic variant of the game.

Kai Arulkumaran, Nat Dilokthanakul, Murray Shanahan et al.

In this paper we combine one method for hierarchical reinforcement learning - the options framework - with deep Q-networks (DQNs) through the use of different "option heads" on the policy network, and a supervisory network for choosing between the different options. We utilise our setup to investigate the effects of architectural constraints in subtasks with positive and negative transfer, across a range of network capacities. We empirically show that our augmented DQN has lower sample complexity when simultaneously learning subtasks with negative transfer, without degrading performance when learning subtasks with positive transfer.