Richard Evans

Piotr Mirowski, Kory W. Mathewson, Jaylen Pittman et al. · deepmind

Language models are increasingly attracting interest from writers. However, such models lack long-range semantic coherence, limiting their usefulness for longform creative writing. We address this limitation by applying language models hierarchically, in a system we call Dramatron. By building structural context via prompt chaining, Dramatron can generate coherent scripts and screenplays complete with title, characters, story beats, location descriptions, and dialogue. We illustrate Dramatron's usefulness as an interactive co-creative system with a user study of 15 theatre and film industry professionals. Participants co-wrote theatre scripts and screenplays with Dramatron and engaged in open-ended interviews. We report critical reflections both from our interviewees and from independent reviewers who watched stagings of the works to illustrate how both Dramatron and hierarchical text generation could be useful for human-machine co-creativity. Finally, we discuss the suitability of Dramatron for co-creativity, ethical considerations -- including plagiarism and bias -- and participatory models for the design and deployment of such tools.

Richard Evans

Medical researchers have solved the problem of estimating the sensitivity and specificity of binary medical diagnostic tests without gold standard tests for comparison. That problem is the same as estimating confusion matrices for classifiers on unlabeled data. This article describes how to modify the diagnostic test solutions to estimate confusion matrices and accuracy statistics for supervised or unsupervised binary classifiers on unlabeled data.

Matthew Shardlow, Richard Evans, Gustavo Henrique Paetzold et al.

This paper presents the results and main findings of SemEval-2021 Task 1 - Lexical Complexity Prediction. We provided participants with an augmented version of the CompLex Corpus (Shardlow et al 2020). CompLex is an English multi-domain corpus in which words and multi-word expressions (MWEs) were annotated with respect to their complexity using a five point Likert scale. SemEval-2021 Task 1 featured two Sub-tasks: Sub-task 1 focused on single words and Sub-task 2 focused on MWEs. The competition attracted 198 teams in total, of which 54 teams submitted official runs on the test data to Sub-task 1 and 37 to Sub-task 2.

Andrew Cropper, Sebastijan Dumančić, Richard Evans et al.

Inductive logic programming (ILP) is a form of logic-based machine learning. The goal is to induce a hypothesis (a logic program) that generalises given training examples. As ILP turns 30, we review the last decade of research. We focus on (i) new meta-level search methods, (ii) techniques for learning recursive programs, (iii) new approaches for predicate invention, and (iv) the use of different technologies. We conclude by discussing current limitations of ILP and directions for future research.

Matthew Shardlow, Richard Evans, Marcos Zampieri

Identifying words which may cause difficulty for a reader is an essential step in most lexical text simplification systems prior to lexical substitution and can also be used for assessing the readability of a text. This task is commonly referred to as Complex Word Identification (CWI) and is often modelled as a supervised classification problem. For training such systems, annotated datasets in which words and sometimes multi-word expressions are labelled regarding complexity are required. In this paper we analyze previous work carried out in this task and investigate the properties of CWI datasets for English. We develop a protocol for the annotation of lexical complexity and use this to annotate a new dataset, CompLex 2.0. We present experiments using both new and old datasets to investigate the nature of lexical complexity. We found that a Likert-scale annotation protocol provides an objective setting that is superior for identifying the complexity of words compared to a binary annotation protocol. We release a new dataset using our new protocol to promote the task of Lexical Complexity Prediction.

Richard Evans, Jose Hernandez-Orallo, Johannes Welbl et al.

The Apperception Engine is an unsupervised learning system. Given a sequence of sensory inputs, it constructs a symbolic causal theory that both explains the sensory sequence and also satisfies a set of unity conditions. The unity conditions insist that the constituents of the theory - objects, properties, and laws - must be integrated into a coherent whole. Once a theory has been constructed, it can be applied to predict future sensor readings, retrodict earlier readings, or impute missing readings. In this paper, we evaluate the Apperception Engine in a diverse variety of domains, including cellular automata, rhythms and simple nursery tunes, multi-modal binding problems, occlusion tasks, and sequence induction intelligence tests. In each domain, we test our engine's ability to predict future sensor values, retrodict earlier sensor values, and impute missing sensory data. The engine performs well in all these domains, significantly outperforming neural net baselines and state of the art inductive logic programming systems. These results are significant because neural nets typically struggle to solve the binding problem (where information from different modalities must somehow be combined together into different aspects of one unified object) and fail to solve occlusion tasks (in which objects are sometimes visible and sometimes obscured from view). We note in particular that in the sequence induction intelligence tests, our system achieved human-level performance. This is notable because our system is not a bespoke system designed specifically to solve intelligence tests, but a general-purpose system that was designed to make sense of any sensory sequence.

Victoria Yaneva, Le An Ha, Richard Evans et al.

When processing a text, humans and machines must disambiguate between different uses of the pronoun it, including non-referential, nominal anaphoric or clause anaphoric ones. In this paper, we use eye-tracking data to learn how humans perform this disambiguation. We use this knowledge to improve the automatic classification of it. We show that by using gaze data and a POS-tagger we are able to significantly outperform a common baseline and classify between three categories of it with an accuracy comparable to that of linguisticbased approaches. In addition, the discriminatory power of specific gaze features informs the way humans process the pronoun, which, to the best of our knowledge, has not been explored using data from a natural reading task.

Richard Evans, Jose Hernandez-Orallo, Johannes Welbl et al.

This paper attempts to answer a central question in unsupervised learning: what does it mean to "make sense" of a sensory sequence? In our formalization, making sense involves constructing a symbolic causal theory that both explains the sensory sequence and also satisfies a set of unity conditions. The unity conditions insist that the constituents of the causal theory -- objects, properties, and laws -- must be integrated into a coherent whole. On our account, making sense of sensory input is a type of program synthesis, but it is unsupervised program synthesis. Our second contribution is a computer implementation, the Apperception Engine, that was designed to satisfy the above requirements. Our system is able to produce interpretable human-readable causal theories from very small amounts of data, because of the strong inductive bias provided by the unity conditions. A causal theory produced by our system is able to predict future sensor readings, as well as retrodict earlier readings, and impute (fill in the blanks of) missing sensory readings, in any combination. We tested the engine in a diverse variety of domains, including cellular automata, rhythms and simple nursery tunes, multi-modal binding problems, occlusion tasks, and sequence induction intelligence tests. In each domain, we test our engine's ability to predict future sensor values, retrodict earlier sensor values, and impute missing sensory data. The engine performs well in all these domains, significantly out-performing neural net baselines. We note in particular that in the sequence induction intelligence tests, our system achieved human-level performance. This is notable because our system is not a bespoke system designed specifically to solve intelligence tests, but a general-purpose system that was designed to make sense of any sensory sequence.

Andrew Cropper, Richard Evans, Mark Law

General game playing (GGP) is a framework for evaluating an agent's general intelligence across a wide range of tasks. In the GGP competition, an agent is given the rules of a game (described as a logic program) that it has never seen before. The task is for the agent to play the game, thus generating game traces. The winner of the GGP competition is the agent that gets the best total score over all the games. In this paper, we invert this task: a learner is given game traces and the task is to learn the rules that could produce the traces. This problem is central to inductive general game playing (IGGP). We introduce a technique that automatically generates IGGP tasks from GGP games. We introduce an IGGP dataset which contains traces from 50 diverse games, such as Sudoku, Sokoban, and Checkers. We claim that IGGP is difficult for existing inductive logic programming (ILP) approaches. To support this claim, we evaluate existing ILP systems on our dataset. Our empirical results show that most of the games cannot be correctly learned by existing systems. The best performing system solves only 40% of the tasks perfectly. Our results suggest that IGGP poses many challenges to existing approaches. Furthermore, because we can automatically generate IGGP tasks from GGP games, our dataset will continue to grow with the GGP competition, as new games are added every year. We therefore think that the IGGP problem and dataset will be valuable for motivating and evaluating future research.

Richard Evans, David Saxton, David Amos et al.

We introduce a new dataset of logical entailments for the purpose of measuring models' ability to capture and exploit the structure of logical expressions against an entailment prediction task. We use this task to compare a series of architectures which are ubiquitous in the sequence-processing literature, in addition to a new model class---PossibleWorldNets---which computes entailment as a "convolution over possible worlds". Results show that convolutional networks present the wrong inductive bias for this class of problems relative to LSTM RNNs, tree-structured neural networks outperform LSTM RNNs due to their enhanced ability to exploit the syntax of logic, and PossibleWorldNets outperform all benchmarks.

Richard Evans, Edward Grefenstette

Artificial Neural Networks are powerful function approximators capable of modelling solutions to a wide variety of problems, both supervised and unsupervised. As their size and expressivity increases, so too does the variance of the model, yielding a nearly ubiquitous overfitting problem. Although mitigated by a variety of model regularisation methods, the common cure is to seek large amounts of training data---which is not necessarily easily obtained---that sufficiently approximates the data distribution of the domain we wish to test on. In contrast, logic programming methods such as Inductive Logic Programming offer an extremely data-efficient process by which models can be trained to reason on symbolic domains. However, these methods are unable to deal with the variety of domains neural networks can be applied to: they are not robust to noise in or mislabelling of inputs, and perhaps more importantly, cannot be applied to non-symbolic domains where the data is ambiguous, such as operating on raw pixels. In this paper, we propose a Differentiable Inductive Logic framework, which can not only solve tasks which traditional ILP systems are suited for, but shows a robustness to noise and error in the training data which ILP cannot cope with. Furthermore, as it is trained by backpropagation against a likelihood objective, it can be hybridised by connecting it with neural networks over ambiguous data in order to be applied to domains which ILP cannot address, while providing data efficiency and generalisation beyond what neural networks on their own can achieve.

Gabriel Dulac-Arnold, Richard Evans, Hado van Hasselt et al.

Being able to reason in an environment with a large number of discrete actions is essential to bringing reinforcement learning to a larger class of problems. Recommender systems, industrial plants and language models are only some of the many real-world tasks involving large numbers of discrete actions for which current methods are difficult or even often impossible to apply. An ability to generalize over the set of actions as well as sub-linear complexity relative to the size of the set are both necessary to handle such tasks. Current approaches are not able to provide both of these, which motivates the work in this paper. Our proposed approach leverages prior information about the actions to embed them in a continuous space upon which it can generalize. Additionally, approximate nearest-neighbor methods allow for logarithmic-time lookup complexity relative to the number of actions, which is necessary for time-wise tractable training. This combined approach allows reinforcement learning methods to be applied to large-scale learning problems previously intractable with current methods. We demonstrate our algorithm's abilities on a series of tasks having up to one million actions.

Peter Sunehag, Richard Evans, Gabriel Dulac-Arnold et al.

Many real-world problems come with action spaces represented as feature vectors. Although high-dimensional control is a largely unsolved problem, there has recently been progress for modest dimensionalities. Here we report on a successful attempt at addressing problems of dimensionality as high as $2000$, of a particular form. Motivated by important applications such as recommendation systems that do not fit the standard reinforcement learning frameworks, we introduce Slate Markov Decision Processes (slate-MDPs). A Slate-MDP is an MDP with a combinatorial action space consisting of slates (tuples) of primitive actions of which one is executed in an underlying MDP. The agent does not control the choice of this executed action and the action might not even be from the slate, e.g., for recommendation systems for which all recommendations can be ignored. We use deep Q-learning based on feature representations of both the state and action to learn the value of whole slates. Unlike existing methods, we optimize for both the combinatorial and sequential aspects of our tasks. The new agent's superiority over agents that either ignore the combinatorial or sequential long-term value aspect is demonstrated on a range of environments with dynamics from a real-world recommendation system. Further, we use deep deterministic policy gradients to learn a policy that for each position of the slate, guides attention towards the part of the action space in which the value is the highest and we only evaluate actions in this area. The attention is used within a sequentially greedy procedure leveraging submodularity. Finally, we show how introducing risk-seeking can dramatically improve the agents performance and ability to discover more far reaching strategies.

Richard Evans

Recent work has shown that dopamine-modulated STDP can solve many of the issues associated with reinforcement learning, such as the distal reward problem. Spiking neural networks provide a useful technique in implementing reinforcement learning in an embodied context as they can deal with continuous parameter spaces and as such are better at generalizing the correct behaviour to perform in a given context. In this project we implement a version of DA-modulated STDP in an embodied robot on a food foraging task. Through simulated dopaminergic neurons we show how the robot is able to learn a sequence of behaviours in order to achieve a food reward. In tests the robot was able to learn food-attraction behaviour, and subsequently unlearn this behaviour when the environment changed, in all 50 trials. Moreover we show that the robot is able to operate in an environment whereby the optimal behaviour changes rapidly and so the agent must constantly relearn. In a more complex environment, consisting of food-containers, the robot was able to learn food-container attraction in 95% of trials, despite the large temporal distance between the correct behaviour and the reward. This is achieved by shifting the dopamine response from the primary stimulus (food) to the secondary stimulus (food-container). Our work provides insights into the reasons behind some observed biological phenomena, such as the bursting behaviour observed in dopaminergic neurons. As well as demonstrating how spiking neural network controlled robots are able to solve a range of reinforcement learning tasks.