Neurophysiologie de la cognition
- Peter Dominey (Stem Cell and Brain Research Institute, INSERM U846, Lyon)
- Laurent Venance (Inserm U667, Collège de France, Paris)
- Hugues Berry (INRIA Rhône-Alpes, Lyon)
|10:30||From the Computational Neuroscience of the Cortico-striatal system to Robot Cognition, par Peter Dominey|
|11:30||The many faces of spike-timing-dependent plasticity: endocannabinoid potentiation underlies fast learning, par Laurent Venance & Hugues Berry|
|12:30||Pause déjeuner : vin et fromage|
From the Computational Neuroscience of the Cortico-striatal system to Robot Cognition
I will present research that began with a computational neuroscience investigation of how the primate cortex and basal ganglia (the Cortico-striatal system in particular) cooperate to form a multimodal associative learning system that has particular impact for temporal sequence learning. This work was extended to include abstract relations that capture aspects of human language processing, and the underlying role of the corticostriatal system. The extended link between language structure and meaning motivated a transition to robotic systems where meaning is incarnated in interaction with the world and other agents. This reaches its climax in the context of human-robot cooperation via shared plans. It has recently been recognized that our initial work in modelling the corticostriatal system in fact helped to lay out the groundwork for the current reservoir computing paradigm, where cortex is the dynamical reservoir and modifiable corticostriatal synapses are the readout mechanism. In a sense we are closing this loop with reservoir computing studies of human language processing.
The many faces of spike-timing-dependent plasticity: endocannabinoid potentiation underlies fast learning
Laurent Venance & Hugues Berry
Synaptic plasticity is the capability of a connection between two neurons to change in strength and is admitted to underlie learning and memory in the brain. However, its experimental demonstration rests on stereotypical experimental protocols that have to be questioned. In particular, the experimental protocols to induce plasticity imply a high number (i.e. hundreds) of stimulations. This holds too for the recently discovered spike-timing-dependent plasticity (STDP) that is usually studied using of the order of one hundred of paired stimulations (PS). Because these requirements seem at odds with the fact that one can learn and recall out of a few trials, we tested whether a STDP could be induced with very low number of PS. Our experimental system is the corticostriatal synapse because of its involvement in procedural learning and memory. We have previously reported a robust STDP at corticostriatal synapses when induced with 100 PS. Here, leveraging mutual interactions between experimental recordings in rat brain slices and a realistic biophysical synapse model, we report dramatic changes in STDP-timing rule when the PS number decreases: (1) 100 to 75 PS induced bidirectional STDP, ie potentiation (increased synaptic strength) or depression (decreased synaptic strength) depending on the timing between stimulations, (2) 50 PS induced unidirectional STDP (only depression is observed) and (3) 10 to 5 PS induced an opposite unidirectional STDP (only potentiation). Note that no significant plasticity was observed for less than 5 PS, denoting a limit of the induction for STDP. Most notably, the biophysical synapse model predicted potentiation (LTP) at 10 to 5 PS to be endocannabinoid-mediated. Strikingly, this prediction was confirmed experimentally: LTP induced by 5-10 post-pre PS was not NMDA-receptor-mediated (unlike the LTP induced by 100 PS) but indeed endocannabinoid-mediated. To test the genericity of this result, we questioned STDP at corticocortical synapses in the somatosensory cortex (between layers 2/3 and 5) and found that we could similarly induce an endocannabinoid-dependent LTP with low PS numbers. Endocannabinoid LTP may thus represent a common form in central structures, suggesting a crucial role for endocannabinoid-mediated STDP in fast learning.