R e s e a r c h _ i n t e r e s t s
I’m interested in reproducing neural learning and memory in computer models.
How are external stimuli acquired into neural systems? How are they retained during time? And can they be used meaningfully later in time?. While neural sciences tend to answer these questions by observing and experimenting on biological networks, new areas of artificial intelligence, such as neuro-evolution, attempt to answer by reproducing neural dynamics in silicon. Neural models take knowledge one step forward in “understanding by doing it” rather than “understanding by observing it”.
Neuromodulation is a key process in selective attention, learning and memory. Whereas excitatory and inhibitory neurons release principally two neurotransmitters, glutamate and GABA, that act by exciting or inhibiting target neurons, neuromodulation involves other chemicals such as Acetylcholine, Dopamine, Norepinephrine and Serotonin that modulate synaptic transmission. Experimental evidence suggests that neuromodulation encodes timed signals to alter synaptic efficacy or plasticity at the occurrence of relevant events such as a surprise or a reward. Modulatory dynamics appears to be an essential element in the implementation of neural learning and memory.
The subject draws inspiration from Neural Science and Computational Neuroscience, Evolutionary Computation, Evolutionary Robotics, Intelligent Control, Adaptive Behaviour, Artificial Life.
E x p e r i m e n t s _ a n d _ c o d e
- Uncertain foraging environments for simulated flying bees
This experiment was used in my CEC 2007 and HIS 2008 papers.
The flight simulation was implemented in C++ without graphical interface. The code is available upon request - Learning agents in T-mazes
The code in C++ is available to reproduce the experiments presented in my ALife XI paper. - Emergent modulatory topologies
Emergent modulatory neurons and topologies are sought in the above enviroments by means of unconstrained evolutionary search on neural topologies.
Highly adaptive and learning networks evolve autonomously to match the requirements in the environments. The C++ code used for the Alife XI and Epigenetic Robotics 2008 papers is available on request. - Analog Genetic Econding
The topology search in my CEC 2007 paper was conducted using the algorithm AGE. For further information on AGE,
please see the website of the Laboratory of Intelligent Systems at EPFL where the algorithm was developed.