Murray Shanahan

ICML Conference 2025 Conference Paper

Learning dynamics in linear recurrent neural networks

• Alexandra Maria Proca
• Clémentine Carla Juliette Dominé
• Murray Shanahan
• Pedro A. M. Mediano

Recurrent neural networks (RNNs) are powerful models used widely in both machine learning and neuroscience to learn tasks with temporal dependencies and to model neural dynamics. However, despite significant advancements in the theory of RNNs, there is still limited understanding of their learning process and the impact of the temporal structure of data. Here, we bridge this gap by analyzing the learning dynamics of linear RNNs (LRNNs) analytically, enabled by a novel framework that accounts for task dynamics. Our mathematical result reveals four key properties of LRNNs: (1) Learning of data singular values is ordered by both scale and temporal precedence, such that singular values that are larger and occur later are learned faster. (2) Task dynamics impact solution stability and extrapolation ability. (3) The loss function contains an effective regularization term that incentivizes small weights and mediates a tradeoff between recurrent and feedforward computation. (4) Recurrence encourages feature learning, as shown through a novel derivation of the neural tangent kernel for finite-width LRNNs. As a final proof-of-concept, we apply our theoretical framework to explain the behavior of LRNNs performing sensory integration tasks. Our work provides a first analytical treatment of the relationship between the temporal dependencies in tasks and learning dynamics in LRNNs, building a foundation for understanding how complex dynamic behavior emerges in cognitive models.

ECAI Conference 2024 Conference Paper

The Propensity for Density in Feed-Forward Models

• Nandi Schoots
• Alex Jackson
• Ali Kholmovia
• Peter McBurney
• Murray Shanahan

Does the process of training a neural network to solve a task tend to use all of the available weights even when the task could be solved with fewer weights? To address this question we study the effects of pruning fully connected, convolutional and residual models while varying their widths. We find that the proportion of weights that can be pruned without degrading performance is largely invariant to model size. Increasing the width of a model has little effect on the density of the pruned model relative to the increase in absolute size of the pruned network. In particular, we find substantial prunability across a large range of model sizes, where our biggest model is 50 times as wide as our smallest model. We explore three hypotheses that could explain these findings. Source code: [29].

TMLR Journal 2023 Journal Article

Feature-Attending Recurrent Modules for Generalization in Reinforcement Learning

• Wilka Torrico Carvalho
• Andrew Kyle Lampinen
• Kyriacos Nikiforou
• Felix Hill
• Murray Shanahan

Many important tasks are defined in terms of object. To generalize across these tasks, a reinforcement learning (RL) agent needs to exploit the structure that the objects induce. Prior work has either hard-coded object-centric features, used complex object-centric generative models, or updated state using local spatial features. However, these approaches have had limited success in enabling general RL agents. Motivated by this, we introduce “Feature- Attending Recurrent Modules” (FARM), an architecture for learning state representations that relies on simple, broadly applicable inductive biases for capturing spatial and temporal regularities. FARM learns a state representation that is distributed across multiple modules that each attend to spatiotemporal features with an expressive feature attention mechanism. We show that this improves an RL agent’s ability to generalize across object-centric tasks. We study task suites in both 2D and 3D environments and find that FARM better generalizes compared to competing architectures that leverage attention or multiple modules.

NeurIPS Conference 2023 Conference Paper

Schema-learning and rebinding as mechanisms of in-context learning and emergence

• Sivaramakrishnan Swaminathan
• Antoine Dedieu
• Rajkumar Vasudeva Raju
• Murray Shanahan
• Miguel Lazaro-Gredilla
• Dileep George

In-context learning (ICL) is one of the most powerful and most unexpected capabilities to emerge in recent transformer-based large language models (LLMs). Yet the mechanisms that underlie it are poorly understood. In this paper, we demonstrate that comparable ICL capabilities can be acquired by an alternative sequence prediction learning method using clone-structured causal graphs (CSCGs). Moreover, a key property of CSCGs is that, unlike transformer-based LLMs, they are {\em interpretable}, which considerably simplifies the task of explaining how ICL works. Specifically, we show that it uses a combination of (a) learning template (schema) circuits for pattern completion, (b) retrieving relevant templates in a context-sensitive manner, and (c) rebinding of novel tokens to appropriate slots in the templates. We go on to marshall evidence for the hypothesis that similar mechanisms underlie ICL in LLMs. For example, we find that, with CSCGs as with LLMs, different capabilities emerge at different levels of overparameterization, suggesting that overparameterization helps in learning more complex template (schema) circuits. By showing how ICL can be achieved with small models and datasets, we open up a path to novel architectures, and take a vital step towards a more general understanding of the mechanics behind this important capability.

ICLR Conference 2023 Conference Paper

Selection-Inference: Exploiting Large Language Models for Interpretable Logical Reasoning

• Antonia Creswell
• Murray Shanahan
• Irina Higgins

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 46 tasks that probe different aspects of logical reasoning. We show that language models tend to perform fairly well at single step inference or entailment tasks, but struggle to chain together multiple reasoning steps to solve more complex problems. In light of this, we propose a Selection-Inference (SI) framework that exploits pre-trained LLMs as general processing modules, and alternates between selection and inference to generate a series of interpretable, casual reasoning steps leading to the final answer. We show that a 7B parameter LLM used within the SI framework in a 5-shot generalisation setting, with no fine-tuning, yields a performance improvement of over 100% compared to an equivalent vanilla baseline on a suite of 10 logical reasoning tasks. The same model in the same setting even outperforms a significantly larger 280B parameter baseline on the same suite of tasks. Moreover, answers produced by the SI framework are accompanied by a causal natural-language-based reasoning trace, which has important implications for the safety and trustworthiness of the system.

IJCAI Conference 2022 Conference Paper

Abstraction for Deep Reinforcement Learning

• Murray Shanahan
• Melanie Mitchell

We characterise the problem of abstraction in the context of deep reinforcement learning. Various well established approaches to analogical reasoning and associative memory might be brought to bear on this issue, but they present difficulties because of the need for end-to-end differentiability. We review developments in AI and machine learning that could facilitate their adoption.

ICLR Conference 2022 Conference Paper

In a Nutshell, the Human Asked for This: Latent Goals for Following Temporal Specifications

• Borja G. León
• Murray Shanahan
• Francesco Belardinelli

We address the problem of building agents whose goal is to learn to execute out-of distribution (OOD) multi-task instructions expressed in temporal logic (TL) by using deep reinforcement learning (DRL). Recent works provided evidence that the agent's neural architecture is a key feature when DRL agents are learning to solve OOD tasks in TL. Yet, the studies on this topic are still in their infancy. In this work, we propose a new deep learning configuration with inductive biases that lead agents to generate latent representations of their current goal, yielding a stronger generalization performance. We use these latent-goal networks within a neuro-symbolic framework that executes multi-task formally-defined instructions and contrast the performance of the proposed neural networks against employing different state-of-the-art (SOTA) architectures when generalizing to unseen instructions in OOD environments.

EWRL Workshop 2022 Workshop Paper

In a Nutshell, the Human Asked for This: Latent Goals for Following Temporal Specifications

• Borja G. León
• Murray Shanahan
• Francesco Belardinelli

We address the problem of building agents whose goal is to learn to execute out-of distribution (OOD) multi-task instructions expressed in temporal logic (TL) by using deep reinforcement learning (DRL). Recent works provided evidence that the agent’s neural architecture is a key feature when DRL agents are learning to solve OOD tasks in TL. Yet, the studies on this topic are still in their infancy. In this work, we propose a new deep learning configuration with inductive biases that lead agents to generate latent representations of their current goal, yielding a stronger generalization performance. We use these latent-goal networks within a neuro-symbolic framework that executes multi-task formallydefined instructions and contrast the performance of the proposed neural networks against employing different state-ofthe-art (SOTA) architectures when generalizing to unseen instructions in OOD environments.

NeurIPS Conference 2021 Conference Paper

Unsupervised Object-Based Transition Models For 3D Partially Observable Environments

• Antonia Creswell
• Rishabh Kabra
• Chris Burgess
• Murray Shanahan

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 transition losses at the level of the object-structured representation rather than pixels. Thanks to the introduction of our novel alignment module, the model deals properly with two issues that are not handled satisfactorily by other transition models, namely object persistence and object identity. We show that the combination of an object-level loss and correct object alignment over time enables the model to outperform a state-of-the-art baseline, and allows it to deal well with object occlusion and re-appearance in partially observable environments.

ICML Conference 2020 Conference Paper

An Explicitly Relational Neural Network Architecture

• Murray Shanahan
• Kyriacos Nikiforou
• Antonia Creswell
• Christos Kaplanis
• David G. T. Barrett
• Marta Garnelo

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 tasks of varying complexity. We show that the proposed architecture, when pre-trained on a curriculum of such tasks, learns to generate reusable representations that better facilitate subsequent learning on previously unseen tasks when compared to a number of baseline architectures. The workings of a successfully trained model are visualised to shed some light on how the architecture functions.

ICML Conference 2020 Conference Paper

Learning to Combine Top-Down and Bottom-Up Signals in Recurrent Neural Networks with Attention over Modules

• Sarthak Mittal
• Alex Lamb
• Anirudh Goyal
• Vikram Voleti
• Murray Shanahan
• Guillaume Lajoie
• Michael Mozer
• Yoshua Bengio

Robust perception relies on both bottom-up and top-down signals. Bottom-up signals consist of what’s directly observed through sensation. Top-down signals consist of beliefs and expectations based on past experience and the current reportable short-term memory, such as how the phrase ‘peanut butter and. .. ’ will be completed. The optimal combination of bottom-up and top-down information remains an open question, but the manner of combination must be dynamic and both context and task dependent. To effectively utilize the wealth of potential top-down information available, and to prevent the cacophony of intermixed signals in a bidirectional architecture, mechanisms are needed to restrict information flow. We explore deep recurrent neural net architectures in which bottom-up and top-down signals are dynamically combined using attention. Modularity of the architecture further restricts the sharing and communication of information. Together, attention and modularity direct information flow, which leads to reliable performance improvements in perceptual and language tasks, and in particular improves robustness to distractions and noisy data. We demonstrate on a variety of benchmarks in language modeling, sequential image classification, video prediction and reinforcement learning that the \emph{bidirectional} information flow can improve results over strong baselines.

ICML Conference 2019 Conference Paper

Policy Consolidation for Continual Reinforcement Learning

• Christos Kaplanis
• Murray Shanahan
• Claudia Clopath

We propose a method for tackling catastrophic forgetting in deep reinforcement learning that is agnostic to the timescale of changes in the distribution of experiences, does not require knowledge of task boundaries and can adapt in continuously changing environments. In our policy consolidation model, the policy network interacts with a cascade of hidden networks that simultaneously remember the agent’s policy at a range of timescales and regularise the current policy by its own history, thereby improving its ability to learn without forgetting. We find that the model improves continual learning relative to baselines on a number of continuous control tasks in single-task, alternating two-task, and multi-agent competitive self-play settings.

ICML Conference 2018 Conference Paper

Conditional Neural Processes

• Marta Garnelo
• Dan Rosenbaum
• Chris J. Maddison
• Tiago Ramalho
• David Saxton
• Murray Shanahan
• Yee Whye Teh
• Danilo Jimenez Rezende

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 appropriate priors. In this paper we propose a family of neural models, Conditional Neural Processes (CNPs), that combine the benefits of both. CNPs are inspired by the flexibility of stochastic processes such as GPs, but are structured as neural networks and trained via gradient descent. CNPs make accurate predictions after observing only a handful of training data points, yet scale to complex functions and large datasets. We demonstrate the performance and versatility of the approach on a range of canonical machine learning tasks, including regression, classification and image completion.

ICML Conference 2018 Conference Paper

Continual Reinforcement Learning with Complex Synapses

• Christos Kaplanis
• Murray Shanahan
• Claudia Clopath

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, an individual synapse in the brain comprises a complex network of interacting biochemical components that evolve at different timescales. In this paper, we show that by equipping tabular and deep reinforcement learning agents with a synaptic model that incorporates this biological complexity (Benna & Fusi, 2016), catastrophic forgetting can be mitigated at multiple timescales. In particular, we find that as well as enabling continual learning across sequential training of two simple tasks, it can also be used to overcome within-task forgetting by reducing the need for an experience replay database.

YNIMG Journal 2015 Journal Article

Effects of lesions on synchrony and metastability in cortical networks

• František Váša
• Murray Shanahan
• Peter J. Hellyer
• Gregory Scott
• Joana Cabral
• Robert Leech

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 Kuramoto model of coupled oscillators with time-delays, calibrated with respect to empirical functional MRI data, to study the relation between the structure of the connectome and two aspects of functional brain dynamics — synchrony, a measure of general coherence, and metastability, a measure of dynamical flexibility. Specifically, we investigate the relationship between the local structure of the connectome, quantified using graph theory, and the synchrony and metastability of the model's dynamics. By removing individual nodes and all of their connections from the model, we study the effect of lesions on both global and local dynamics. Of the nine nodal graph-theoretical properties tested, two were able to predict effects of node lesion on the global dynamics. The removal of nodes with high eigenvector centrality leads to decreases in global synchrony and increases in global metastability, as does the removal of hub nodes joining topologically segregated network modules. At the level of local dynamics in the neighbourhood of the lesioned node, structural properties of the lesioned nodes hold more predictive power, as five nodal graph theoretical measures are related to changes in local dynamics following node lesions. We discuss these results in the context of empirical studies of stroke and functional brain dynamics.

ECAI Conference 2006 Conference Paper

Leaf Classification Using Navigation-Based Skeletons

• Georgios Sakellariou
• Murray Shanahan

In this paper, we present a leaf classification method based on skeletons produced by a navigation-inspired technique. The classification system comprises three separate stages. First, a skeletonisation algorithm is used to gather low level structural and morphological information about the shape. Subsequently, the data is converted into a series of attributed graphs. Graphs of the same type are then compared using an approximate graph matcher, which identifies a degree of similarity between them. Each degree of similarity corresponds to a dimension in a conceptual space, as defined by Gärdenfors. We test the performance of our technique on a set of leaves belonging to three different species.

ECAI Conference 2006 Conference Paper

On Packing 2D Irregular Shapes

• Alexandros Bouganis
• Murray Shanahan

Designing and implementing an intelligent system that tackles the problem of placing two-dimensional shapes on a surface, such that no shapes overlap and the uncovered surface area is minimized, is highly important in industrial applications. However, it is also interesting from the scientific perspective, in terms of artificial intelligence, since autonomous systems developed up to now have found it difficult to compete with humans in this task. This paper presents a new algorithm which addresses the on-line packing of two-dimensional irregular shapes, and achieves high quality solutions in short computational times. The key point of this algorithm is the utilization of techniques drawn from computer vision and artificial intelligence.

KR Conference 2004 Conference Paper

A Logic-Based Formulation of Active Visual Perception

• David Randell
• Murray Shanahan

Building on earlier attempts to characterise robot perception as a form of abduction, this paper presents a logical account of active visual perception in the context of an upper-torso humanoid robot. Using the event calculus to represent actions and change, and aspect graphs to represent shape, the formalisation captures the way the robot’s knowledge of the objects in its workspace can increase through interaction with them.

AIJ Journal 2004 Journal Article

An attempt to formalise a non-trivial benchmark problem in common sense reasoning

• Murray Shanahan

Most logic-based AI research works at a meta-theoretical level, producing new logics and studying their properties. Little effort is made to show how these logics can be used to formalise object-level theories of common sense. In the spirit of Pat Hayes's Naive Physics Manifesto, the present paper supplies a formalisation of a non-trivial benchmark problem in common sense physical reasoning, namely how to crack an egg. The formalisation is based on the event calculus, a well-known formalism for reasoning about action. Along the way, a number of methodological issues are raised, such as the question of how the symbols deployed in the formalisation might be grounded through a robot's interaction with the world.

ECAI Conference 2004 Conference Paper

Indirect and Conditional Sensing in the Event Calculus

• Jeremy Forth
• Murray Shanahan

IJCAI Conference 2003 Conference Paper

A Logic-based Algorithm for Image Sequence Interpretation and Anchoring

• Paulo Santos
• Murray Shanahan

This paper describes a logic-based framework for interpretation of sequences of scenes captured by a stereo vision system of a mobile robot. An algorithm for anchoring and interpretation of such sequences is also proposed.

KR Conference 2002 Conference Paper

A Logical Account of Perception Incorporating Feedback and Expectation

• Murray Shanahan

ECAI Conference 2002 Conference Paper

Hypothesising Object Relations from Image Transitions

• Paulo E. Santos
• Murray Shanahan

AIJ Journal 2000 Journal Article

M. Shanahan, Solving the Frame Problem

• Murray Shanahan

IJCAI Conference 1999 Conference Paper

The Ramification Problem in the Event Calculus

• Murray Shanahan

Finding a solution to the frame problem that is robust in the presence of actions with indirect effects has proven to be a difficult task. Examples that feature the instantaneous propagation of interacting indirect effects are particularly taxing. This article shows that an already widely known predicate calculus formalism, namely the event calculus, can handle such examples with only minor enhancements.

AAAI Conference 1997 Conference Paper

Noise, Non-Determinism and Spatial Uncertainty

• Murray Shanahan

This paper presents a logical account of sensor data assimilation in a mobile robot, based on abduction. Unlike previous work, the present formulation handles sensor noise as well as motor noise. In addition, it incorporates two significant technical advances. The use of determining fluents to deal with non-determinism obviates the need for a special form of abduction, and the use of uncertain object boundaries alleviates a problem with multiple explanations.

AAAI Conference 1996 Conference Paper

Noise and the Common Sense Informatic Situation for a Mobile Robot

• Murray Shanahan

Any model of the world a robot constructs on the basis of its sensor data is necessarily both incomplete, due to the robot’ s limited window on the world, and uncertain, due to sensor and motor noise. This paper supplies a logical account of sensor data assimilation in which such models are constructed through an abductive process which hypothesises the existence, locations, and shapes of objects. Noise is treated as a kind of non-determinism, and is dealt with by a consistency-based form of abduction.

ECAI Conference 1996 Conference Paper

Robotics and the Common Sense Informatic Situation

• Murray Shanahan

AIJ Journal 1995 Journal Article

A circumscriptive calculus of events

• Murray Shanahan

A calculus of events is presented in which domain constraints, concurrent events, and events with nondeterministic effects can be represented. The paper offers a nonmonotonic solution to the frame problem for this formalism that combines two of the techniques developed for the situation calculus, namely causal and state-based minimisation. A theorem is presented which guarantees that temporal projection will not interfere with minimisation in this solution, even in domains with ramifications, concurrency, and nondeterminism. Finally, the paper shows how the formalism can be extended to cope with continuous change, whilst preserving the conditions for the theorem to apply.

AIJ Journal 1995 Journal Article

Default reasoning about spatial occupancy

• Murray Shanahan

This paper describes a default reasoning problem, analogous to the frame problem, that arises when an attempt is made to construct a logic-based calculus for reasoning about the movement of objects in a real-valued co-ordinate system. A number of potential solutions to this problem are examined. Particular attention is given to the interaction between the default reasoning required by these solutions and that required to overcome the frame problem, especially when the latter demands an “existence of situations” axiom.

IJCAI Conference 1993 Conference Paper

Explanation in the Situation Calculus

• Murray Shanahan

This paper explores different techniques for explanation within the framework of the situation calculus, using the so-called stolen car problem as its main example. Two approaches to explanation are compared: the deductive approach usually found in the literature, and a less common abductive approach. Both approaches are studied in the context of two different styles of representation.

ECAI Conference 1990 Conference Paper

Representing Continuous Change in the Event Calculus

• Murray Shanahan

IJCAI Conference 1989 Conference Paper

Prediction is Deduction but Explanation is Abduction

• Murray Shanahan

IJCAI Conference 1987 Conference Paper

An Incremental Theorem Prover

• Murray Shanahan