Murray Shanahan

Murray Shanahan
Imperial College London | Imperial · Department of Computing

PhD

About

173
Publications
64,443
Reads
How we measure 'reads'
A 'read' is counted each time someone views a publication summary (such as the title, abstract, and list of authors), clicks on a figure, or views or downloads the full-text. Learn more
6,609
Citations
Citations since 2016
46 Research Items
3123 Citations
20162017201820192020202120220100200300400500
20162017201820192020202120220100200300400500
20162017201820192020202120220100200300400500
20162017201820192020202120220100200300400500

Publications

Publications (173)
Preprint
Full-text available
Although contemporary large language models (LMs) demonstrate impressive question-answering capabilities, their answers are typically the product of a single call to the model. This entails an unwelcome degree of opacity and compromises performance, especially on problems that are inherently multi-step. To address these limitations, we show how LMs...
Article
Human perception and experience of time are strongly influenced by ongoing stimulation, memory of past experiences, and required task context. When paying attention to time, time experience seems to expand; when distracted, it seems to contract. When considering time based on memory, the experience may be different than what is in the moment, exemp...
Preprint
Full-text available
Large language models (LLMs) have been shown to be capable of impressive few-shot generalisation to new tasks. However, they still tend to perform poorly on multi-step logical reasoning problems. Here we carry out a comprehensive evaluation of LLMs on 50 tasks that probe different aspects of logical reasoning. We show that language models tend to p...
Article
Full-text available
Artificial Intelligence is making rapid and remarkable progress in the development of more sophisticated and powerful systems. However, the acknowledgement of several problems with modern machine learning approaches has prompted a shift in AI benchmarking away from task-oriented testing (such as Chess and Go) towards ability-oriented testing, in wh...
Article
Full-text available
The apparent dichotomy between information-processing and dynamical approaches to complexity science forces researchers to choose between two diverging sets of tools and explanations, creating conflict and often hindering scientific progress. Nonetheless, given the shared theoretical goals between both approaches, it is reasonable to conjecture the...
Preprint
Full-text available
The apparent dichotomy between information-processing and dynamical approaches to complexity science forces researchers to choose between two diverging sets of tools and explanations, creating conflict and often hindering scientific progress. Nonetheless, given the shared theoretical goals between both approaches, it is reasonable to conjecture the...
Preprint
Full-text available
We present an architecture that is effective for continual learning in an especially demanding setting, where task boundaries do not exist or are unknown. Our architecture comprises an encoder, pre-trained on a separate dataset, and an ensemble of simple one-layer classifiers. Two main innovations are required to make this combination work. First,...
Preprint
Artificial Intelligence is making rapid and remarkable progress in the development of more sophisticated and powerful systems. However, the acknowledgement of several problems with modern machine learning approaches has prompted a shift in AI benchmarking away from task-oriented testing (such as Chess and Go) towards ability-oriented testing, in wh...
Preprint
We present a slot-wise, object-based transition model that decomposes a scene into objects, aligns them (with respect to a slot-wise object memory) to maintain a consistent order across time, and predicts how those objects evolve over successive frames. The model is trained end-to-end without supervision using losses at the level of the object-stru...
Article
Full-text available
The problem of common sense remains a major obstacle to progress in artificial intelligence. Here, we argue that common sense in humans is founded on a set of basic capacities that are possessed by many other animals, capacities pertaining to the understanding of objects, space, and causality. The field of animal cognition has developed numerous ex...
Preprint
Recently developed deep learning models are able to learn to segment scenes into component objects without supervision. This opens many new and exciting avenues of research, allowing agents to take objects (or entities) as inputs, rather that pixels. Unfortunately, while these models provide excellent segmentation of a single frame, they do not kee...
Preprint
Full-text available
In this paper, we propose a multi-timescale replay (MTR) buffer for improving continual learning in RL agents faced with environments that are changing continuously over time at timescales that are unknown to the agent. The basic MTR buffer comprises a cascade of sub-buffers that accumulate experiences at different timescales, enabling the agent to...
Preprint
Full-text available
Human perception and experience of time is strongly influenced by ongoing stimulation, memory of past experiences, and required task context. When paying attention to time, time experience seems to expand; when distracted, it seems to contract. When considering time based on memory, the experience may be different than in the moment, exemplified by...
Article
Full-text available
In the history of the quest for human-level artificial intelligence, a number of rival paradigms have vied for supremacy. Symbolic artificial intelligence was dominant for much of the 20th century, but currently a connectionist paradigm is in the ascendant, namely machine learning with deep neural networks. However, both paradigms have strengths an...
Preprint
Full-text available
Recent advances in artificial intelligence have been strongly driven by the use of game environments for training and evaluating agents. Games are often accessible and versatile, with well-defined state-transitions and goals allowing for intensive training and experimentation. However, agents trained in a particular environment are usually tested o...
Preprint
With a view to bridging the gap between deep learning and symbolic AI, we present a novel end-to-end neural network architecture that learns to form propositional representations with an explicitly relational structure from raw pixel data. In order to evaluate and analyse the architecture, we introduce a family of simple visual relational reasoning...
Preprint
We propose a method for tackling catastrophic forgetting in deep reinforcement learning that is \textit{agnostic} to the timescale of changes in the distribution of experiences, does not require knowledge of task boundaries, and can adapt in \textit{continuously} changing environments. In our \textit{policy consolidation} model, the policy network...
Article
Full-text available
Despite being a fundamental dimension of experience, how the human brain generates the perception of time remains unknown. Here, we provide a novel explanation for how human time perception might be accomplished, based on non-temporal perceptual classification processes. To demonstrate this proposal, we build an artificial neural system centred on...
Article
Full-text available
[This corrects the article DOI: 10.1371/journal.pone.0189109.].
Chapter
In most contemporary work in deep reinforcement learning (DRL), agents are trained in simulated environments. Not only are simulated environments fast and inexpensive, they are also ‘safe’. By contrast, training in a real world environment (using robots, for example) is not only slow and costly, but actions can also result in irreversible damage, e...
Preprint
Stochastic video prediction is usually framed as an extrapolation problem where the goal is to sample a sequence of consecutive future image frames conditioned on a sequence of observed past frames. For the most part, algorithms for this task generate future video frames sequentially in an autoregressive fashion, which is slow and requires the inpu...
Preprint
Deep neural networks excel at function approximation, yet they are typically trained from scratch for each new function. On the other hand, Bayesian methods, such as Gaussian Processes (GPs), exploit prior knowledge to quickly infer the shape of a new function at test time. Yet GPs are computationally expensive, and it can be hard to design appropr...
Preprint
Full-text available
We introduce an approach for deep reinforcement learning (RL) that improves upon the efficiency, generalization capacity, and interpretability of conventional approaches through structured perception and relational reasoning. It uses self-attention to iteratively reason about the relations between entities in a scene and to guide a model-free polic...
Preprint
Despite being a fundamental dimension of experience, how the human brain generates the perception of time remains unknown. Here, we provide a novel explanation for how human time perception might be accomplished, based on non-temporal perceptual clas-sification processes. To demonstrate this proposal, we built an artificial neural system centred on...
Article
Unlike humans, who are capable of continual learning over their lifetimes, artificial neural networks have long been known to suffer from a phenomenon known as catastrophic forgetting, whereby new learning can lead to abrupt erasure of previously acquired knowledge. Whereas in a neural network the parameters are typically modelled as scalar values,...
Article
Full-text available
Although brain oscillations involving the basal ganglia (BG) have been the target of extensive research, the main focus lies disproportionally on oscillations generated within the BG circuit rather than other sources, such as cortical areas. We remedy this here by investigating the influence of various cortical frequency bands on the intrinsic effe...
Data
Impact of the cortical phase offset to the TE of GPe and STN in different frequency ranges. Inner circle: Normalized TE of GPe (A) and STN (B) afferents versus efferents across phase offsets ϕ. Outer circle: Normalized TE of the indirect (A), hyper-direct (B) and direct (C) pathways as defined in Fig 5. (EPS)
Data
Effect of short-term plasticity in synaptic conductances. Ratio between steady-state conductance G and the initial value G0 for different pre-synaptic spike frequencies, and for all plastic connections of the BG circuit. (EPS)
Data
Connectivity of the STN in the phasic mode. The STN sends diffuse connections to the GPe that spread across all simulated neighbouring channels. The transparency of the arrows represents the firing rate of the source structure. (EPS)
Data
Transfer entropy for various delays. Animated visualization of transfer entropy between the BG structures and the cortex for different values of the time delay of information flow. SD1/2: MSN neurons with d1/2 dopamine receptors respectively, T2: Phasically-active cortical ensemble. (GIF)
Preprint
Full-text available
A plethora of evidence and theoretical work indicates that the basal ganglia (BG) might be the locus where conflicts between prospective motor programs, or actions, are being resolved. Similarly to the majority of brain regions, oscillations in this subcortical group are ubiquitous, largely driven by the cortex and associated with a number of motor...
Article
Full-text available
In this work we study the distributed representations learnt by generative neural network models. In particular, we investigate the properties of redundant and synergistic information that groups of hidden neurons contain about the target variable. To this end, we use an emerging branch of information theory called partial information decomposition...
Article
While information processing in complex systems can be described in abstract, general terms, there are cases in which the relation between these computations and the physical substrate of the underlying system is itself of interest. Prominently, the brain is one such case. With the aim of relating information and dynamics in biological neural syste...
Article
Full-text available
Continuous-time recurrent neural networks are widely used as models of neural dynamics and also have applications in machine learning. But their dynamics are not yet well understood, especially when they are driven by external stimuli. In this article, we study the response of stable and unstable networks to different harmonically oscillating stimu...
Article
The problem of sparse rewards is one of the hardest challenges in contemporary reinforcement learning. Hierarchical reinforcement learning (HRL) tackles this problem by using a set of temporally-extended actions, or options, each of which has its own subgoal. These subgoals are normally handcrafted for specific tasks. Here, though, we introduce a g...
Article
Full-text available
We study a variant of the variational autoencoder model with a Gaussian mixture as a prior distribution, with the goal of performing unsupervised clustering through deep generative models. We observe that the standard variational approach in these models is unsuited for unsupervised clustering, and mitigate this problem by leveraging a principled i...
Article
Full-text available
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 lea...
Article
Full-text available
It has been shown that sets of oscillators in a modular network can exhibit a rich variety of metastable chimera states, in which synchronisation and desynchronisation coexist. Independently, under the guise of integrated information theory, researchers have attempted to quantify the extent to which a complex dynamical system presents a balance of...
Article
Deep reinforcement learning is the learning of multiple levels of hierarchical representations for reinforcement learning. Hierarchical reinforcement learning focuses on temporal abstractions in planning and learning, allowing temporally-extended actions to be transferred between tasks. In this paper we combine one method for hierarchical reinforce...
Chapter
It has been demonstrated that, in a network of excitatory and inhibitory neurons, a synchronous response gradually emerges due to spike timing dependant plasticity acting upon an external spatio-temporal stimulus that is repeatedly applied. This paper builds on these findings by addressing two questions relating to STDP and network dynamics. Firstl...
Article
Current theory proposes that healthy neural dynamics operate in a metastable regime, where brain regions interact to simultaneously maximize integration and segregation. Metastability may confer important behavioral properties, such as cognitive flexibility. It is increasingly recognized that neural dynamics are constrained by the underlying struct...
Article
Full-text available
At the macroscopic scale, the human brain can be described as a complex network of white matter tracts integrating grey matter assemblies - the human connectome. The structure of the connectome, which is often described using graph theoretic approaches, can be used to model macroscopic brain function at low computational cost. Here, we use the Kura...
Article
This paper critically assesses the anti-functionalist stance on consciousness adopted by certain advocates of integrated information theory (IIT), a corollary of which is that human-level artificial intelligence implemented on conventional computing hardware is necessarily not conscious. The critique draws on variations of a well-known gradual neur...
Conference Paper
Neuronal avalanches are a cortical phenomenon characterised by bursts of activity bracketed by periods of quiescence. It has been shown both in vivo and in vitro that the size and length of avalanche events conform to power law-like distributions, suggesting the system is within or near a critical state. This work investigates the interplay of netw...
Conference Paper
Low-frequency oscillations have been the target of extensive research both in cortical structures and in the basal ganglia, due to numerous reports of associations with brain disorders and the normal functioning of the brain. Whereas a number of computational models of the basal ganglia investigate these phenomena, these models tend to focus on int...
Article
Full-text available
Entropy is a dimensionless quantity that is used for measuring uncertainty about the state of a system but it can also imply physical qualities, where high entropy is synonymous with high disorder. Entropy is applied here in the context of states of consciousness and their associated neural dynamics, with a particular focus on the psychedelic state...
Article
Full-text available
Understanding how dynamic changes in brain activity control behavior is a major challenge of cognitive neuroscience. Here, we consider the brain as a complex dynamic system and define two measures of brain dynamics: the synchrony of brain activity, measured by the spatial coherence of the BOLD signal across regions of the brain; and metastability,...
Conference Paper
Full-text available
It has been proposed that oscillating groups of neurons firing synchronously provide a mechanism that underlies many cognitive functions. In previous work, it has been demonstrated that, in a network of excitatory and inhibitory neurons, a synchronous response gradually emerges due to spike timing dependant plasticity (STDP) acting upon an external...
Conference Paper
This work proposes a biologically plausible cognitive architecture implemented in spiking neurons, which is based on well- established models of neuronal global workspace, action selection in the basal ganglia and corticothalamic circuits and can be used to control agents in virtual or physical environments. The aim of this system is the investigat...
Article
Full-text available
Many species of birds, including pigeons, possess demonstrable cognitive capacities, and some are capable of cognitive feats matching those of apes. Since mammalian cortex is laminar while the avian telencephalon is nucleated, it is natural to ask whether the brains of these two cognitively capable taxa, despite their apparent anatomical dissimilar...
Article
Full-text available
Groups of neurons firing synchronously are hypothesized to underlie many cognitive functions such as attention, associative learning, memory, and sensory selection. Recent theories suggest that transient periods of synchronization and desynchronization provide a mechanism for dynamically integrating and forming coalitions of functionally related ne...
Data
Average number of correlations, and the peak of modulated exploration scatter plot. This plot is of the original 250 data points from which the surface plot of figure 6 was created. The setup is the same as for figure 3 and subsequent figures. (A) The average number of mean intermittent frequency correlations found normalised by the number of coexi...
Data
Supporting Figure Descriptions. DOC)
Data
Mean intermittent frequency correlation scatter plot. This plot is of the original 250 data points from which the surface plot of figure 5 was created. The setup is the same as for figure 3 and subsequent figures. (A) The mean intermittent frequency correlation for networks using the QIF neuron model, (B) the same as panel A for the HH neuron model...
Data
Separation of positive and anti mean intermittent frequency correlation. (A) The positive mean intermittent frequency correlation for the QIF neuron model. (B) The same as panel A for the HH neuron model. (C) The anti mean intermittent frequency correlation for the QIF neuron model. (B) The same as panel C for the HH neuron model. (TIF)
Article
This paper presents a toolbox of solutions that enable the user to construct biologically-inspired spiking neural networks with tens of thousands of neurons and millions of connections that can be simulated in real time, visualized in 3D and connected to robots and other devices. NeMo is a high performance simulator that works with a variety of neu...
Article
Full-text available
Modular networks of delay-coupled and pulse-coupled oscillators are presented, which display both transient (metastable) synchronization dynamics and the formation of a large number of "chimera" states characterized by coexistent synchronized and desynchronized subsystems. We consider networks based on both community and small-world topologies. It...
Article
Full-text available
This paper addresses the question of how the brain of an animal achieves cognitive integration--that is to say how it manages to bring its fullest resources to bear on an ongoing situation. To fully exploit its cognitive resources, whether inherited or acquired through experience, it must be possible for unanticipated coalitions of brain processes...
Conference Paper
Full-text available
Background / Purpose: This presentation shows recent work in which we extracted the large-scale “wiring diagram” of the pigeon forebrain. Although the brain of a pigeon is anatomically very different from the brain of a primate, they turn out to be organised along similar lines in terms of connectivity. Main conclusion: Both are modular, small...
Article
According to the singularity hypothesis, rapid and accelerating technological progress will in due course lead to the creation of a human-level artificial intelligence capable of designing a successor artificial intelligence of significantly greater cognitive prowess, and this will inaugurate a series of increasingly super-intelligent machines. But...
Conference Paper
Full-text available
A novel clustering algorithm is presented for analyzing the temporal dynamics of synchronization in networks of coupled oscillators and applied to a model of resting-state brain activity. Connectivity in the model is based on a human-brain structural connectivity matrix derived from diffusion tensor imaging tractography. We find a strong correspond...