I am a postdoctoral student in the Cognitive NeuroRobotics Unit at the Okinawa Institute of Science and Technology with Jun Tani, since January 2018. My work focuses on developmental processes in robots. The goal is to further approach open-ended, cumulative, lifelong learning in robots.
This postdoc follows a previous one in the Mnemosyne team at Inria and at the Institute of Neurodegenerative Diseases in Bordeaux, France, with Nicolas Rougier. I did my my PhD in developmental robotics with Pierre-Yves Oudeyer, where I focused my research on issues of exploration, learning and behavioral diversity.
My research investigates developmental processes in robots. Much of the research efforts in machine learning and robotics are currently conducted toward developing learning algorithms to improve performance at specific tasks. This does not solve the problem of open-ended lifelong learning, where a robot would be expected to discover and learn diverse knowledge and skills through autonomous and social learning in the real world, at lifelong timescales. Such a learning process would require, among other things, developmental processes embedded into the robot itself to guide its discovery of a complex world, much like the one that are at play in infants and children.
I am currently collaborating with Randall O’Reilly’s Computational Cognitive Neuroscience Laboratory from the University of Colorado Boulder. The goal of the collaboration is a reimplementation of the computation neuroscience framework Leabra in Python that is quantitatively equivalent to the one in the emergent software. In parallel, I work with the Jonathan Cohen’s team at the Princeton Neuroscience Institute to integrate the implementation into the PsyNeuLink framework.
I am committed to the produce open and reproducible science. All my publications are freely available and come along with their code, released under the Open Science License. I am involved in several replication efforts and I am a reviewer at the ReScience journal.
The Recode project aims at reimplementing experiments described in published scientific articles. Besides replicating published results, those implementations strive to be as short, as simple and as understandable as possible, so that all the code can be shown alongside the explanations.