Complexity-based AI
This research line exploits the interface between Complexity Science methods and concepts (statistical physics, dynamical systems, time series, ...) and AI (mainly neural network models for supervised and unsupervised learning). We use complexity techniques to unravel the inner workings of neural network training optimization, and we get inspired by the physics of complexity to build decentralized AI solutions where the whole is more than the sum of its parts.
Key papers on theory of machine learning
ANN training through the lens of a dynamicist
Kaloyan Danovski, Miguel C. Soriano and Lucas Lacasa
Frontiers in Complex Systems (in press)
More is different: interacting brains and collective learning
An effective theory of collective deep learning
Lluís Arola-Fernández and Lucas Lacasa
Submitted for publication