Periodic Reporting for period 1 - SpikeControl (Cerebellar Spiking Model For Real-time Closed-loop Sensorimotor Control)
Reporting period: 2015-09-01 to 2017-08-31
The VOR depends on the vestibular system, which detects head rotation. VOR’s nature is purely feed-forward since it induces prompt compensatory eye movements as consequence of head movements. VOR is mediated by a control system in which adaptation is directly driven by sensorimotor errors: the cerebellum. The existing mismatch between head movements (signalled by the vestibular organ) and the incoming information to the cerebellum about eye movements represents sensory errors, which are called retinal slips. The feed-forward adaptive control mediated by the cerebellum aims at minimising these retinal slips. The VOR, together with eye-blink classical conditioning, is broadly assumed as the paradigm that better reveals cerebellar learning.
During SPIKECONTROL, we modelled the neural basis of VOR control to provide a mechanistic understanding of the cerebellar functionality, which plays a key role in VOR adaptation. This work focused in testing the main theoretical hypotheses through this modelling approach. On the one hand, this work aimed at cross-linking data on VOR at behavioural and neural level. On the other hand, the developed VOR controller, based on cerebellar sensorimotor adaptation, was integrated within the simulated iCub and the actual iCub robot. Through the simulation of VOR control impairments, we examined possible consequences on the vestibular processing capabilities of the VOR model
First, we focused in the inferior olive–Purkinje cell–cerebellar nuclei IO–PC–CN subcircuit, thus the granular layer was initially simplified. We tested the hypothesis of the CN afferents operating as a gain-controller and a learning consolidation site thanks to the Spike-timing dependent-plasticity mechanisms (STDP) underlying the cerebellar neural structure. STDP operates not only at parallel fibres PF–PC synapses but also at mossy fibres MF–CN and PC–CN afferents. We demonstrated how these STDP mechanisms contributed to create an internal adaptive gain controller at CN afferents whilst the synaptic distribution at PF–PC synapses were transferred to MF– CN synapses
Second, the focus was in the cerebellar cortex, concretely the granular layer. We studied how to evaluate the input activity within the cerebellar granule-cell layer. Supervised learning has long been attributed to several feed forward neural circuits within the brain, paying particular attention to the cerebellar granule-cell layer amongst them. The activity of cerebellar granule cells is conveyed by PF and translated into PC activity.We developed an algorithm for evaluating quantitatively the level of compatibility/interference amongst a set of given cerebellar states according to their representation (granule-cell activation patterns) without the need for actually conducting simulations and network training. We configured the granular layer activations according to the given output of this algorithm and then interconnected this layer to the IO-PC-CN sub circuit, thus creating the neural basis for our cerebellar model that was used in subsequence project stages
Third, once the cerebellar model working principles were tested, we further developed the cerebellar model as a VOR controller. This controller was embedded in a forward control loop that allowed cerebellar-dependent sensorimotor adaptation in a neuromimetic system
Finally, the cerebellar capability of performing adaptive information processing mediating sensorimotor control was evaluated in a specific VOR test through a closed-loop sensorimotor interaction in real time (RT). Simulating our cerebellar-based VOR model conjointly with the humanoid robotic platform helped assessing both qualitative and quantitative performance indicators to constructively cross-validate the model against neurophysiological and behavioural data. To that aim the cerebellar-based VOR model were implemented and tested first in the iCub simulator, and finally in the real humanoid iCub robot. The VOR controller worked in RT as the humanoid robot iCub moved its head
The cerebellum self-adapts better than artificial systems to manipulation tasks with changeable dynamics. Despite the tremendous progress in computational neuroscience, electronics, and mechatronics being made, building a synthetic controller able to outperform a human being when moving and interacting with others remains being a major challenge. Interestingly, this is true despite the enormous information processing capacity, in theory, that the 3rd generation of artificial neural networks holds (Spiking neural network). This strongly suggests that advances in embodied neuroscience will depend on removing current technical limitations, thus bringing spiking neural networks to ‘intelligent machines’ applications. To that aim, we developed within this project one amongst the first neuromorphic architectures that can run medium scale cerebellar simulations able to operate an iCub in RT