Skip to main content

Cerebellar Distributed Plasticity Towards Active Sensing and Motor Control

Periodic Reporting for period 1 - CEREBSENSING (Cerebellar Distributed Plasticity Towards Active Sensing and Motor Control)

Reporting period: 2015-11-01 to 2017-10-31

Although robotic system capabilities have experienced a remarkable boost in the last decades, their movements still remain far from looking dexterous. If state-of-the-art robotic systems could operate outside the highly constrained environments they are currently restricted to, a wide-variety of applications in manufacturing, medicine, elderly support and general domestic applications would open up. A crucial capability is still missing: the ability to perceive, understand and discriminate relevant information in complex own movement realizations. Active proprioception in humans takes advantage of previously stored dynamics models of movements and efferent copies of the motor commands to predict the sensorial consequences of the actions, outperforming in this way the passive-sensing counterpart. Interestingly, the cerebellar forward model suggests that this sensorial prediction may enhance the motor control by avoiding the effect of the sensorial feedback delay (estimated around 100ms).

According to traditional theories, cerebellum plasticity is driven by a teaching signal reaching the Purkinje cells (PC) through the climbing fibers (CF) from the Inferior Olive (IO). This model has provided the basis for solving simple associative tasks such as eye-blink conditioning or ocular reflexes and more complex manipulation tasks. However, two experimental discoveries have recently questioned this simple model: (i) late anatomical studies have shown that cerebellar granular layer cells (granule -GrCs- and Golgi -GoCs- ) receive multimodal convergent connections carrying separate sensory (proprioceptive) and motor-related information (supporting the cerebellar role in sensory processing) and, (ii) synaptic sites in the cerebellar granular layer and deep cerebellar nuclei (DCN) (additionally to the originally proposed PF-PC) have shown traces of plasticity. Computational models have emerged as a powerful tool in order to explain the role of these additional plasticity sites.

CEREBSENSING aims to understand how the cerebellum processes sensorial information coming from the cerebral cortex by using computational models embedded in realistic perception-action simulations. More specifically, this project has created computational models of the cerebellum including plasticity at the cerebellar granular layer. Our simulations suggest that plasticity at the inhibitory interneuron afferents (namely, the Golgi cells) is effective in creating sparse representations of the input information. The enhanced sensorial representation at the granular layer provides the basis for forward sensorial consequence estimation at the subsequent layer (Purkinje cells) as it emerges from the integration of the granular layer in a whole-cerebellum model controlling the saccade movements. Finally, the application of the granular layer structure into a visual digit classification task has evidenced how this neuronal network can be applied for real-life applications.
During the development of the CEREBSENSING project two different simulation frameworks for cerebellar models have been implemented:
- The High-Performance Computing (HPC) simulation framework allows the simulation of large-scale cerebellar models in distributed memory supercomputers. This system is convenient for theoretical study of information transmission with previously generated input and offline simulation data analysis and it was used for cerebellar granular layer simulation in WP2.
- The real-time closed-loop framework for robotic control allows the simulation of medium-scale cerebellar models in real-time in personal computers. It interfaces an efficient neural network simulator (EDLUT) with the Robotic Operating System (ROS) and the physical or simulated robotic systems (namely, the Baxter humanoid robot and the eye plant).

The first simulation framework is at the base of the implementation and analysis of a large-scale model of the plasticity at the granular layer. This model reproduces the neuronal distribution and synaptic connection of a 1mm-size die and includes STDP plasticity at every synaptic layer. The Golgi cell plasticity (excitatory and inhibitory) has shown especially relevant in creating sparse representations of the input information.The granular layer model has been embedded in a whole-cerebellum model and it has been simulated with the real-time close-loop framework. The cerebellar model successes on controlling the eye movements in a saccade movement paradigm. After the learning process, the cerebellar model accurately predicts the eye movement at the Purkinje cell activity (as proposed for the forward predictor theory of the cerebellum) and generates precise motor commands at the fastigial nucleus (compatible with the inverse model theory of the cerebellum)

Finally, regarding the industrial application of the granular layer information processing system, The granular-layer inspired network model has been applied for a digit recognition task. It is based on the MNIST-DVS dataset. This dataset contains the events produced by a dynamic vision sensor (DVS) during the recording of handwritten digits in the MNIST dataset. The network, including only unsupervised plasticity, has shown able to recognize the digits been presented in the input.

The fellow has been notably active in disseminating the results emerging from the project within the scientific community (mainly in neuroscience, robotics, neuromorphic computing and artificial intelligence). During the two year project duration, the fellow has co-authored 7 journal articles in top journals. He has contributed to 8 international conferences or workshops (including the most prestigious conferences in cerebellum and computational neuroscience) and has taught 3 invited talks.
The human cerebellum presents astonishing adaptation capabilities outperforming the most advanced artificial robotic controllers. It can be observed in the dexterous movements that human being present after a short learning process of training in new tasks. Despite the impressive progress in computing performance and machine learning methods (including deep learning techniques) it does not seem to have remarkable impact on movement control of robotic systems and even the most impressive systems from Boston Dynamics still remains far from the movements that a kid would perform. Hence, this strongly suggests that advances in robotics will depend on increasing current knowledge on brain operation (and especially on the cerebellum). To that aim, we have established a new theory of how cerebellum plasticity may enhance sensorial information processing at the input layer.

Providing embodied neurosystems able to take into account changeable dynamics and, therefore, able to interact with human beings opens up an entire new economic and social scenario in health and care industries, especially in rehabilitation. We could envision new robotic devices that would drive the rehabilitation process of patients. These robots may adapt its operation to the evolution of the patient. Similarly, these cerebellum models could be used for controlling exoskeletons in many different conditions. For example, exercising old people for improving life quality when they are confined to bed, or lifting them out of their beds in order to do cleaning and to eliminate bedsores. Robot assisted lifting can potentially extend the dignity of elderly people as well as support care workers.
Summary description of the learning mechanisms involved in cerebellar adaptation