Cognition seminars

Social world models extend prediction beyond physical dynamics to the intentions, beliefs and strategies of interacting agents. This talk examines how such representations can be learned and used in socially capable AI. Examples include multimodal models of behaviour, benchmarks for social reasoning and agents that adapt over extended interactions.

Human recall of meaningful narratives suggests that comprehension builds a hierarchical memory representation. This talk presents the recall evidence, a mathematical model reproducing the observed organization, and language-model-assisted semantic segmentation supporting a related hierarchical account of language.

Infants develop concepts and relations before acquiring substantial language, often with little explicit supervision. This talk presents computational models of that early learning and compares them with language-centred AI training. It asks whether the conceptual structures formed before language could provide useful ingredients for artificial learning systems.

Dopamine-based reinforcement is well studied in artificial tasks, but its role in learning natural behaviour needs direct evidence. This talk tracks juvenile zebra-finch song alongside dopamine signals in a song-related basal-ganglia region. Signals distinguish renditions approaching the eventual adult song from poorer versions, predict later behavioural change and depend on recent performance history. These observations support a prediction-error account of reinforcement during developmental vocal learning.

Weaver ants coordinate their bodies and surrounding leaves to construct enclosed nests. Laboratory experiments show that a local zipping rule predicts construction in simple settings, yet colonies also solve puzzles where that rule fails. The talk examines how coupled ant–leaf dynamics can support information processing beyond individual cognition, providing an experimental model for the capabilities and limits of collective problem solving.

Behaviours that resemble world knowledge do not by themselves reveal what an AI represents internally. This talk uses Craik's account of mental models—supporting alternatives, prediction and transfer—to frame empirical tests. Spatial reasoning in language models and motion forecasting in visual models illustrate ways to probe those abilities, along with the limits of inferring world models from systems trained on human-produced representations.