Emergence, Self-Organisation, and the Future of Adaptive AI

by José Sousa (AI4 Continuous Learning Lab (AI4CL) Multidisciplinary Institute of Ageing (MIA) University of Coimbra)

Wednesday 8 Apr 2026, 14:00 → 15:00 Portugal

Sala de Conferências (Departamento de Física FCTUC)

The dominant paradigm in Artificial Intelligence (AI) rests on a single premise: that learning scales with data and computation. Large Language Models (LLMs) and Deep Learning (DL) architectures have normalised this assumption, yet they remain fundamentally brittle — unable to generalise beyond their training distribution, incapable of continuous learning, and dependent on human feedback to adapt. This talk proposes a different foundation.
Drawing on the physics of complex adaptive systems, we present GALILEAN, a project developing a theory and architecture for AI continuous learning grounded in adaptive spiking neural networks (aSNNs). The central hypothesis is that tacit knowledge — experiential, non-codified, and resistant to explicit representation — can emerge as a collective property of interacting intelligent agents, much as macroscopic order emerges from microscopic interactions in statistical physics.
We model this as an aSNN in which agents form probabilistic beliefs from raw data without training or validation pipelines, share those beliefs through spike trains, and self-organise from fully connected to scale-free network topologies driven by spike-timing-dependent plasticity. Evolutionary game theory (EGT) provides the measurement language, mapping agent interaction dynamics onto attractor landscapes in phase space.
The framework is empirically grounded in a longitudinal clinical dataset involving nine clinicians diagnosing allergy diseases, with further validation planned for age-related macular degeneration diagnosis. Open questions — on the conditions for emergence, the role of information coupling, and the stability of knowledge attractors — are framed across physics, network science, cognitive sciences and information theory.

https://www.uc.pt/mia/research/ai-4-continuous-learning/