SocialAI Research Group

SocialAI

SocialAI develops multi-agent neural network models grounded in cutting edge approaches from neuroscience, psychology, sociology, history, and cultural anthropology. We aim to illuminate the cognitive processes that enable people to navigate fundamental aspects of cooperation and competition within and between social groups.

Research directions

How and why do humans act the way they do in groups? How do behaviours like stereotyping, internalized social roles, labour specialization, and prosociality emerge from the repetitive interaction of multiple agents?

How can we unpack the complex set of interactions that exist at different levels of analysis – from cognitive processes to small-group behaviours to the norms and institutions expressed by whole societies?

How can we build better, more biologically-grounded and empirically-informed AI models, and how can these help us navigate the future of social interactions, both virtual and real-world?

Biologically-inspired and cognitively plausible AI systems

Developing machine-learning processes based in, and built around, current understanding of complex cognitive processes, including around memory formation and retrieval, affective emotional states, and biases in social interaction

GEM: Breakthrough MARL environment building

In-house bespoke software for multiagent RL, Gem, offers a novel approach to environment building and experiment design, addressing key design questions: (1) how can our experiment setup be most inviting to social scientists with minimal background in machine learning? (2) what setup allows a complex environments and behaviours to emerge from agent behaviour? (3) how can our experiments serve as plausible, empirically-grounded abstractions of real world phenomena?

Understand – and predict – emergent behaviours

By tracing how the effects AI system design choices and the interaction of behaviour across levels of analysis produce various outcomes, our frameworks will help us explore how this will impact future social behaviour

Multi-level investigation from neuron to nation-state

We can’t hope to understand individual choices, group behaviour, or societal outcomes without considering all levels of analysis. Our models integrate empirically-informed insights across these levels.

SocialAI Research Group 2023333