Building global models from local circuits
Task leader: ETIS.
There is a computational problem dealing with spiking neurons for controlling robots or software agents in real time. But even more important is that it is still unclear what kind of information is used by neurons, groups of neurons or brain structures in order to process information. Tasks 1, 2 and 3 will shed a light on the relevant information: first spike, medial spike, phase, mean firing rate, synchronization, etc. Moreover, it may be that first spike information is important at the neuron level, but synchronization at the structure level. Hence a global model may only be understood as arising from local circuits. Following this idea, it is important to model the spiking properties in collaboration with GNT and to see to what extend these properties have to be replaced or kept depending on the task to perform. Hence our goal is not to build a different architecture for each neurobiological experiment, but rather to come up with one single architecture dealing with all experiments.
The ETIS hippocampus-prefrontal cortex model (where future choices are coded as place-to-place transitions) will be integrated with the GNT spiking model, in consultation with LPPA and LNC for neurophysiologically based algorithms.
Robotic implementation of these models in challenges based upon the behavioral neuroscience experiments will permit comparison of the relative benefits and drawbacks of the respective modeling approaches in approaching (or surpassing) the performance of the biological systems.
Subtask 1 - Modelling the generation of prospective coding in hippocampus and prefrontal cortex
We will examine the generation of prospective activity within the hippocampal loop, particularly in the entorhinal cortex layer 3 recurrent network.
Subtask 2 - Building sequences
The second subtask concerns improving the biological plausibility (and hence the robustness and flexibility) of our Hpc-Pfc system model (Figure 2)
Sketch of the ETIS neurocybernetics Hippocampal and prefrontal model.
Currently there is a one-to-one connectivity between transition coding neurons of the Hpc CA1 area (Fig. 3) and the nodes of the cognitive map of the Pfc. The identical coding in these two structures must be updated on the basis of the data of our neuroscience partners.
Cognitive map built in an open area comparable to a rodent experimental setup.
Subtask 3 - Action selection
ETIS will add the Pfc-striatal results (subtask 2 of task 3) in the architecture defined in subtasks 1 and 2 of task 4. It will also use the results provided by subtask 3 of task 3. This should allow us to test the robot’s ability to learn to navigate towards goal sites whether by conditioning or by the use of cognitive maps, or perhaps a combination of the two, depending upon learning level. This capacity permits it to remain functional even in the absence of one of the signal processing pathways, as seen in animals (for instance going to the goal in the dark).