Research interests

We want to understand how to extract knowledge from data, from data combined with previous knowledge, or from a combination of different pieces of knowledge.

We develop new mathematical formalisms, adapt existing ones, and apply them to exciting problems like collective behavior and processing in biological circuits.

Keywords: Model Theory, Abstract Algebra, Reinforcement Learning, Deep Learning, abstraction, collective behavior, Neuroscience, zebrafish, biological circuits, dolphin communication


Awarded FCT grant to apply AI methods to predict and understand behavior (2021-2024)

Awarded H2020 ALMA to continue developing algebraic AI (2020-2024)

Awarded H2020 FindingPheno to apply ML to multi-omics data (2021-2024)

Helping CrowdFight COVID-19

5 days lectures at EACA School of Computer Algebra

Latest results

We are building machine intelligence systems that are transparent to mathematical analysis. We also use these systems to study problems in collective behavior.

Martin-Maroto F, de Polavieja GG. Algebraic Machine Learning.
[ ArXiv ]
We propose an approach to learning using Abtract Algebra and Model Theory (and not any optimization) that treats data and formal knowledge in the same manner. All requirements of the system are written as sentences in a formal language and a model in which the sentences are valid in the found. This formalization helps in building theorems to show in detail how these systems learn. FUTURE: We are working on the formal study of these type of learning systems, and we will soon release software to play with this approach. The H2020 project ALMA (ALgebraic MAchines) is designed to go deeper into its foundations, to make easier its interaction with humans and to build some concrete applications.

In the application to collective behavior, we build sytems that extract information from raw data and models that are predictive and insightful.

Romero-Ferrero F, Bergomi MG, Hinz RC, Heras FJ, de Polavieja GG. idtracker. ai: tracking all individuals in small or large collectives of unmarked animals. Nature Methods. 2019 Feb;16(2):179-82.
[ journal ] [ pdf ] [ web ] [ gitlab ]
Building on our idea of tracking by identification, see idTracker, here we propose to do the identification using deep nets to scale better with the number of animals. One network self-learns from a video which image blobs are individuals and which crossings. A second network self-learns to distinguish each individual. It is a general purpose open source system of lab animal behavior in any species, and we use it to acquire data of collective fish behavior. FUTURE: check its website for updates adapting the system for a better user-experience and improving the internal methods.

Heras FJ, Romero-Ferrero F, Hinz RC, de Polavieja GG. Deep attention networks reveal the rules of collective motion in zebrafish. PLoS computational biology. 2019 Sep 13;15(9):e1007354.
[ journal ] [ biorxiv ] [ gitlab ]
We use the power of deep nets to have a predictive model of collective behavior. To avoid the black-box nature of deep nets, we propose a concrete modular structure for the network that makes the model understandable. This structure implies the modelling assumption that interactions among agents are in pairs and the decision is based on a weighted mean of these pair interactions. We checked that, when applied to artificial data obtained of agents moving according to some mathematical rules, the method recovers the underlying rules. When applied to data of fish collective behavior obtained with, we learn that each fish dynamically focuses on different subgroups of other fish depending not only on where they are but also on what direction and speed they have. FUTURE: We are applying these models to more controlled experiments to zoom into some properties of collective behavior.

Costa T, Laan A, Heras FJH, de Polavieja GG. Automated discovery of local rules for desired collective-level behavior through reinforcement learning. Fundamentals and Applications of AI: An Interdisciplinary Perspective, Front. Phys. 8: 200. doi: 10.3389/fphy, 2020.
[ journal ] [ gitlab ] [ videos ]
We set out to obtain animal interactions that can explain beautiful global structures of collective behavior seen in Nature like rotating balls, rotating tornadoes and rotating mills. We model each individual with a minimal cognitive apparatus (with or without a simple model of the retina) and moving in a medium of modelled Physics. Reinforcement Learning is used to find the sensorimotor transformations that results in the observed 3D structure. We found they are strikingly similar to the ones we see in the lab, with differences mainly in the z-direction to explain the different 3D structures. FUTURE: We are studying the advantages of this type of global models in concert with more local (trajectory) models.


Gonzalo G. de Polavieja (PI),
[ scholar ] [ ORCID CV ] [ Faculty web ] [ twitter ]

Fernando Martin-Maroto (Senior Researcher),
[ Faculty web ] [ twitter ]

Francisco J. H. Heras (postdoc),
[ scholar ] [ ORCID CV ] [ Faculty web ] [ twitter ]

David Mendez (postdoc),
[ scholar ] [ ORCID CV ] [ twitter ]

Francisco Romero-Ferrero (PhD student),
[ scholar ] [ ORCID CV ] [ Faculty web ] [ twitter ]

Dean Rance (PhD student),
[ twitter ]

Useful internal contacts

Lab manager (ordering): Telma Carrilho,

Human Resources (contract, card, e-mail address): Teresa Carona,

Pre-award (grant and fellowship applications): Andreia Tavares,

Post-award (management of awarded grants): Vanda Vicente,

Operations Manager (general): Catia Feliciano,