Basic problems in learning from data and collective behavior
Mathematical modelling and analysis tools, AI, behavior
Algebras, Zebrafish and human groups, general data structures
We study basic problems in acquiring knowledge from data.
Keywords: Collective behavior, machine learning, AI, compression, algebraic embedding
Compression and learning:
We are pursuing a principled approach to Machine Learning to be able to derive some basic properties about learning.
We derived a relationship between compression and accuracy. We also derived which types of compression are the ones leading to learning.
We are also pursuing a parameter-free approach in which we use training data to build a small algebraic model. We believe this approach might have the seeds to learning concepts from data,
How to obtain good models of collective behavior:
We have obtained principled approaches to collective animal behavior based on estimation and control theory. These start from a mathematization of a simple hypothesis and derive relationships we test in zebrafish and other experiments, including data from humans.
We are complementing this approach with a data-driven approach based on deep learning models of collective behavior using versions of interaction and attention networks and reinforcement learning.
To obtain good datasets of collective behavior, we also build tracking systems, our latest being idtracker.ai
If you are interested in basic problems of learning from data or collective behavior, like maths and coding, work both independently and in a team, like to understand well a problem and also present it nicely to the community, be open about code an data, maybe this is your place.
Contact us at email@example.com
Visit also our Champalimaud site
Champalimaud Centre for the Unknown
Avenida Brasília, 1400-038 Lisbon, Portugal
T (+351) 210 480 200