This webpage provides access to presentations given during the Scientific Discussion Meeting, with the remaining materials available from here.
- Professor Leslie Valiant FRS, Harvard University, USA
Educability
[slides]
Abstract:
We seek to define the capability that has enabled humans to develop the civilisation we have, and that distinguishes us from other species. For this it is not enough to identify a distinguishing characteristic - we want a capability that is also explanatory of humanity's achievements. "Intelligence" does not work here because we have no agreed definition of what intelligence is or how an intelligent entity behaves. We need a concept that is behaviourally better defined. The definition will need to be computational in the sense that the expected outcomes of exercising the capability need to be both specifiable and computationally feasible. This formulation is related to the goals of AI research but is not synonymous with it, leaving out the many capabilities we share with other species.
We make a proposal for this essential human capability, which we call "educability." It synthesises abilities to learn from experience, to learn from others, and to chain together what we have learned in either mode and apply that to particular situations. It starts with the now standard notion of learning from examples, as captured by the Probably Approximately Correct model and used in machine learning. The ability of Large Language Models learning from examples to generate smoothly flowing prose lends encouragement to this approach. The basic question then is how to extend this approach to encompass broader human capabilities beyond learning from examples. This is what the educability notion aims to answer.
- Professor Luc De Raedt, KU Leuven and Örebro University, Belgium
How to make logics neurosymbolic
[slides]
Abstract:
Neurosymbolic AI (NeSy) is regarded as the third wave in AI. It aims at combining knowledge representation and reasoning with neural networks. Numerous approaches to NeSy are being developed and there exists an `alphabet-soup' of different systems, whose relationships are often unclear. I will discuss the state-of-the art in NeSy and argue that there are many similarities with statistical relational AI (StarAI).
Taking inspiring from StarAI, and exploiting these similarities, I will argue that Neurosymbolic AI = Logic + Probability + Neural Networks. I will also provide a recipe for developing NeSy approaches: start from a logic, add a probabilistic interpretation, and then turn neural networks into `neural predicates'.
Probability is interpreted broadly here, and is necessary to provide a quantitative and differentiable component to the logic. At the semantic and the computation level, one can then combine logical circuits (aka proof structures) labelled with probability, and neural networks in computation graphs.
I will illustrate the recipe with NeSy systems such as DeepProbLog, a deep probabilistic extension of Prolog, and DeepStochLog, a neural network extension of stochastic definite clause grammars (or stochastic logic programs).
- Dr Jessica Hamrick, Google DeepMind, UK
Planning, reasoning, and generalisation in deep learning
[slides]
Abstract:
What do we need to build artificial agents which can reason effectively and generalise to new situations? An oft-cited claim, both in cognitive science and in machine learning, is that a key ingredient for reasoning and generalisation is planning with a model of the world. In this talk, Dr Hamrick will evaluate this claim in the context of model-based reinforcement learning, presenting evidence that demonstrates the utility of planning for certain classes of problems (e.g. in-distribution learning and procedural generalisation in reinforcement learning), as well as evidence that planning is not a silver bullet for out-of-distribution generalisation. In particular, generalisation performance is limited by the generalisation abilities of the individual components required for planning (e.g., the policy, reward model, and world model), which in turn are dependent on the diversity of data those components are trained on. Moreover, generalisation is strongly dependent on choosing the appropriate level of abstraction. These concerns may be partially addressed by leveraging new state-of-the-art foundation models, which are trained on both an unprecedented breadth of data and at a higher level of abstraction than before.
- Professor Leslie Pack Kaelbling, Massachusetts Institute of Technology, USA
The role of rationality in modern AI
[slides]
Abstract:
The classical approach to AI was to design systems that were rational at run-time: they had explicit representations of beliefs, goals, and plans and ran inference algorithms, online, to select actions. The rational approach was criticised (by the behaviourists) and modified (by the probabilists) but persisted in some form. Now the overwhelming success of the connectionist approach in so many areas presents evidence that the rational view may no longer have a role to play in AI. This talk examines this question from several perspectives, including whether the rationality is present at design-time and/or at run-time, and whether systems with run-time rationality might be useful from the perspectives of computational efficiency, cognitive modelling and safety. It will present some current research focused on understanding the roles of learning in runtime-rational systems with the ultimate aim of constructing general-purpose human-level intelligent robots.
- Professor Timothy Behrens FRS, University of Oxford and University College London, UK
Representing the future
Abstract:
To flexibly adapt to new situations, our brains must understand the regularities in the world, but also in our own patterns of behaviour. A wealth of findings is beginning to reveal the algorithms we use to map the outside world. In contrast, the biological algorithms that map the complex structured behaviours we compose to reach our goals remain enigmatic. Here we reveal a neuronal implementation of an algorithm for mapping abstract behavioural structure and transferring it to new scenarios. We trained mice on many tasks which shared a common structure organising a sequence of goals but differed in the specific goal locations. Animals discovered the underlying task structure, enabling zero-shot inferences on the first trial of new tasks. The activity of most neurons in the medial Frontal cortex tiled progress-to-goal, akin to how place cells map physical space. These “goal-progress cells” generalised, stretching and compressing their tiling to accommodate different goal distances. In contrast, progress along the overall sequence of goals was not encoded explicitly. Instead, a subset of goal-progress cells was further tuned such that individual neurons fired with a fixed task-lag from a particular behavioural step. Together these cells implemented an algorithm that instantaneously encoded the entire sequence of future behavioural steps, and whose dynamics automatically retrieved the appropriate action at each step. These dynamics mirrored the abstract task structure both on-task and during offline sleep. Our findings suggest that goal-progress cells in the medial frontal cortex may be elemental building blocks of schemata that can be sculpted to represent complex behavioural structures.
- Professor Evelina Fedorenko, Massachusetts Institute of Technology, USA
Language is distinct from thought in the human brain
[slides]
Abstract:
Dr Fedorenko seeks to understand how humans understand and produce language, and how language relates to, and works together with, the rest of human cognition. She will discuss the ‘core’ language network, which includes left-hemisphere frontal and temporal areas, and show that this network is ubiquitously engaged during language processing across input and output modalities, strongly interconnected, and causally important for language. This language network plausibly stores language knowledge and supports linguistic computations related to accessing words and constructions from memory and combining them to interpret (decode) or generate (encode) linguistic messages. Importantly, the language network is sharply distinct from higher-level systems of knowledge and reasoning. First, the language areas show little neural activity when individuals solve math problems, infer patterns from data, or reason about others’ minds. And second, some individuals with severe aphasia lose the ability to understand and produce language but can still do math, play chess, and reason about the world. Thus, language does not appear to be necessary for thinking and reasoning. Human thinking instead relies on several brain systems, including the network that supports social reasoning and the network that supports abstract formal reasoning. These systems are sometimes engaged during language use—and thus have to work with the language system—but are not language-selective. Many exciting questions remain about the representations and computations in the systems of thought and about how the language system interacts with these higher-level systems. Furthermore, the sharp separation between language and thought in the human brain has implications for how we think about this relationship in the context of AI models, and for what we can expect from neural network models trained solely on linguistic input with the next-word prediction objective.
- Professor Terrence W Deacon, University of California, USA
Neither nature nor nurture: the semiotic infrastructure of symbolic reference
[slides]
Abstract:
Use of the symbol concept suffers from a conflation of two interpretations: a) a conventional sign vehicle (an alphanumeric character), and b) a conventional reference relationship (word meaning). Both are often mischaracterised in terms of "arbitrarity," a negative attribute. When your computer begins randomly displaying characters (a) on your screen, they are generally interpreted as indications of malfunction (or the operation of a "viral" algorithm). And yet when LLMs print out strings of characters that are iconic of interpretable sentences, we assume (b) that they are more than mere icons and indices of an algorithmic (aka mechanistic) process. This begs the question of what distinguishes symbolic interpretation from iconic and indexical interpretation and how they are related. Conventional relations are not just "given," however, they must be acquired. As a result, they are dependent on prior non-conventional referential relations (i.e. iconic and indexical interpretive processes) to extrinsically "ground" the reference of these intrinsically "ungrounded" sign vehicles. This semiotic infrastructure exemplifies the hierarchic complexity of symbolic reference, why it is cognitively difficult for non-humans, and hints at the special neurological architecture that aids human symbolic cognition. It also is relevant for understanding the difference between how humans and generative AI systems produce and process the tokens used as sign vehicles. So far, LLMs and their generative cousins are structured by token-token iconic and indexical relations only (though of extremely high dimensionality), not externally grounded by iconic and indexical pragmatic relations, even though the token-token relations of the training data have been.
- Professor Emma Brunskill, Stanford University, USA
Learning to make decisions from few examples
[slides]
Abstract:
Humans have the ability to quickly learn new tasks, but many machine learning algorithms require enormous amounts of data to perform well. One critical arena of tasks involves decision making under uncertainty, learning from data to make good decisions to optimise expected utility. In this talk I’ll consider this challenge from a computational lens, discussing algorithms that help reveal when learning to make good decisions is easy or hard, algorithmic approaches that can change the order of how many data points are required, and multi-task learning algorithms that can automatically infer and leverage structure across tasks to substantially improve performance.
- Professor Yee Whye Teh, University of Oxford and Google DeepMind, UK
Do we even need to learn representations in machine learning?
Abstract:
Learning representations from raw data signals is a long standing goal of machine learning. This is underlined by the fact that one of the major conferences in the area is called International Conference on Learning Representations (ICLR). Good representations are expected to enable generalisation, compositionality and interpretability, and to serve as a bridge between the observed raw data signals and the abstract world of concepts and symbols. Representation learning is particularly important to the field of deep generative models, which has historically aimed to learn latent variables which effectively represent raw signals, as well as encoders and decoders which map between latent variables and raw signals. Recent advances in generative AI, such as language models and diffusion models, has put deep generative models in the limelight. However he will argue that these advances have been achieved by effectively giving up on the goal of representation learning. This begs the question of whether representation learning is a mirage, and what are we missing on the road to understanding intelligence?
- Professor Richard S Sutton FRS, University of Alberta, Canada
Dynamic deep learning
[slides]
Abstract:
Deep learning and large language models have dramatically shifted the conversation about Signals vs Symbols in favour of numerical methods. Nevertheless, current deep learning methods are limited; they have great difficulty learning during their normal operation. In this talk, Sutton argues that this is not an inherent limitation of neural networks, but just of the algorithms currently used, and he proposes new neural-network algorithms specifically designed for continual learning.
- Professor Sheila McIlraith, University of Toronto and Vector Institute, Canada
How (formal) language can help AI agents learn, plan, and remember
[slides]
Abstract:
Humans have evolved languages over tens of thousands of years to provide useful abstractions for understanding and interacting with each other and with the physical world. Language comes in many forms. In Computer Science and in the study of AI, we have historically used knowledge representation languages and programming languages to capture our understanding of the world and to communicate unambiguously with computers. In this talk I will discuss how (formal) language can help agents learn, plan, and remember in the context of reinforcement learning. I’ll show how we can exploit the compositional syntax and semantics of formal language and automata to aid in the specification of complex reward-worthy behaviour, to improve the sample efficiency of learning, and to help agents learn what is necessary to remember. In doing so, I argue that (formal) language can help us address some of the challenges to reinforcement learning in the real world.
- Professor Alexandre Pouget, University of Geneva, Switzerland
Neural models of compositional learning
[slides]
Abstract:
Compositionality is widely regarded as a key component of general intelligence, yet its neural basis has remained elusive, largely due to the absence of plausible neural models. Recently, however, advancements in large language models (LLMs) and a resurgence of interest in recurrent neural networks (RNNs) have led to the development of models that exhibit compositional behaviours across various tasks. In this presentation, we introduce two such models. The first leverages LLMs to instruct RNNs to execute a task based solely on natural language instructions. The second model demonstrates how complex motor behaviours can be decomposed into sequences of simpler motor primitives. In both cases, adopting a compositional approach significantly reduces the learning time for tasks, even enabling 0-shot learning in scenarios where traditional reinforcement learning (RL) algorithms would require thousands or millions of training iterations. With these neural models, we now have the tools to experimentally test various hypotheses about compositionality in the brain, both in humans and animals.
- Dr Jane X Wang, Google DeepMind, UK
Meta-learning as bridging the neuro-symbolic gap in AI
[slides]
Abstract:
Meta-learning, the ability to acquire and utilise prior knowledge to facilitate new learning, is a hallmark of human and animal cognition. This capability is also evident in deep reinforcement learning agents and large language models (LLMs). While recurrent neural networks have offered insights into the neural underpinnings of learning, understanding how LLMs, trained on vast amounts of human-generated text and code, achieve rapid in-context learning remains a challenge. The inherent structure present in these training data sources—reflecting symbolic knowledge embedded in language and cultural artifacts—potentially plays a crucial role in enabling LLMs to generalise effectively. Therefore, examining the role of structured data in large-scale model training through a cognitive science lens offers crucial insights into how these models acquire and generalise knowledge, mirroring aspects of human learning.
This talk will discuss how these findings not only deepen our understanding of deep learning models but also offer potential avenues for integrating machine learning and symbolic reasoning, through the lens of meta-learning and cognitive science. Insights from meta-learning research can inform the development of embodied AI agents, such as those in the Scalable, Instructable, Multiworld Agent (SIMA) project, by incorporating structured knowledge representations and meta-learning capabilities to potentially enhance their ability to follow instructions, generalise to novel tasks, and interact more effectively within many complex 3D environments.
- Professor Christopher Summerfield, University of Oxford and UK AI Safety Institute, UK
The Habermas Machine: using AI to help people find common ground
[slides]
Abstract:
Language models allow us to treat text as data. This opens up new opportunities for human communication, deliberation, and debate. I will describe a project in which we use an LLM to help people find agreement, by training it to produce statements about political issues that a group with diverse views will endorse. We find that the statements it produces help people find common ground, and shift their views towards a shared stance on the issue. By analysing embeddings, we show that the group statements respect the majority view but prominently include dissenting voices. We use the tool to mount a virtual citizens’ assembly and show that independent groups debating political issues relevant to the UK move in a common direction. We call this AI system the “Habermas Machine”, after the theorist Jurgen Habermas, who proposed that when rational people debate under idealised conditions, agreement will emerge in the public sphere.
- Professor Moritz Hardt, Max Planck Institute of Intelligent Systems, Germany
The emerging science of benchmarks
[slides]
Abstract:
Benchmarks have played a central role in the progress of machine learning research since the 1980s. Although there's much researchers have done with them, we still know little about how and why benchmarks work. In this talk, I will trace the rudiments of an emerging science of benchmarks through selected empirical and theoretical observations. Looking back at the ImageNet era, I'll discuss what we learned about the validity of model rankings and the role of label errors. Looking ahead, I'll talk about new challenges to benchmarking and evaluation in the era of large language models. The results we'll encounter challenge conventional wisdom and underscore the benefits of developing a science of benchmarks.
