Limits...
A bi-hemispheric neuronal network model of the cerebellum with spontaneous climbing fiber firing produces asymmetrical motor learning during robot control.

Pinzon-Morales RD, Hirata Y - Front Neural Circuits (2014)

Bottom Line: The bi-hemispheric structure is inspired by the observation that lesioning one hemisphere reduces motor performance asymmetrically.Our results showed that asymmetrical conditions were successfully handled by the biCNN model, in contrast to a single hemisphere model or a classical non-adaptive proportional and derivative controller.Thus, we conclude that a bi-hemispheric structure and adequate spontaneous activity of cf inputs are critical for cerebellar asymmetrical motor learning.

View Article: PubMed Central - PubMed

Affiliation: Neural Cybernetics Laboratory, Department of Computer Science, Chubu University Kasugai, Japan.

ABSTRACT
To acquire and maintain precise movement controls over a lifespan, changes in the physical and physiological characteristics of muscles must be compensated for adaptively. The cerebellum plays a crucial role in such adaptation. Changes in muscle characteristics are not always symmetrical. For example, it is unlikely that muscles that bend and straighten a joint will change to the same degree. Thus, different (i.e., asymmetrical) adaptation is required for bending and straightening motions. To date, little is known about the role of the cerebellum in asymmetrical adaptation. Here, we investigate the cerebellar mechanisms required for asymmetrical adaptation using a bi-hemispheric cerebellar neuronal network model (biCNN). The bi-hemispheric structure is inspired by the observation that lesioning one hemisphere reduces motor performance asymmetrically. The biCNN model was constructed to run in real-time and used to control an unstable two-wheeled balancing robot. The load of the robot and its environment were modified to create asymmetrical perturbations. Plasticity at parallel fiber-Purkinje cell synapses in the biCNN model was driven by error signal in the climbing fiber (cf) input. This cf input was configured to increase and decrease its firing rate from its spontaneous firing rate (approximately 1 Hz) with sensory errors in the preferred and non-preferred direction of each hemisphere, as demonstrated in the monkey cerebellum. Our results showed that asymmetrical conditions were successfully handled by the biCNN model, in contrast to a single hemisphere model or a classical non-adaptive proportional and derivative controller. Further, the spontaneous activity of the cf, while relatively small, was critical for balancing the contribution of each cerebellar hemisphere to the overall motor command sent to the robot. Eliminating the spontaneous activity compromised the asymmetrical learning capabilities of the biCNN model. Thus, we conclude that a bi-hemispheric structure and adequate spontaneous activity of cf inputs are critical for cerebellar asymmetrical motor learning.

Show MeSH

Related in: MedlinePlus

Address Event Representation (AER) for implementation of the biCNN model. (A) Example of AER for a network with four neurons. (B) Vectors required for representing the network in (A) by using AER. The specific case for neuron #3 is shown in red. N stands for the number of neurons (N = 4 in the example), and Nsyn, the number of synapses in the network (Nsyn = 6 in the example).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4221029&req=5

Figure 4: Address Event Representation (AER) for implementation of the biCNN model. (A) Example of AER for a network with four neurons. (B) Vectors required for representing the network in (A) by using AER. The specific case for neuron #3 is shown in red. N stands for the number of neurons (N = 4 in the example), and Nsyn, the number of synapses in the network (Nsyn = 6 in the example).

Mentions: Address Event Representation (AER) is a communication technique for sparse networks and has been successfully extrapolated to neural networks (Johansson and Lansner, 2007). In AER, four vectors are required to describe the network architecture (Figure 4B). The first vector ID encodes the neurons in the network, assigning each a unique ID. The second vector NP stores the number of pre-synapses for each neuron, and the third and fourth vectors P and W encode the IDs of the pre-synaptic neurons and the corresponding synaptic weights in a stacked, ordered way. For example, in Figure 4A, neuron #3 is contacted by two neurons (#2, #1); this information is clearly observed in the third element of vectors ID[2] = 3, NP[2] = 2 (red arrows Figure 4B). By accumulating the number of synapses for the neurons that precede the neuron #3 (neuron #2, 2 pre-synapses, neuron #1, 1 pre-synapse), the index for the pre-synapses of neuron #3 can be read in the vectors P[3] = 1, and P[4] = 2, with their respective weights in W[3] = 0.3, and W[4] = −0.4.


A bi-hemispheric neuronal network model of the cerebellum with spontaneous climbing fiber firing produces asymmetrical motor learning during robot control.

Pinzon-Morales RD, Hirata Y - Front Neural Circuits (2014)

Address Event Representation (AER) for implementation of the biCNN model. (A) Example of AER for a network with four neurons. (B) Vectors required for representing the network in (A) by using AER. The specific case for neuron #3 is shown in red. N stands for the number of neurons (N = 4 in the example), and Nsyn, the number of synapses in the network (Nsyn = 6 in the example).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4221029&req=5

Figure 4: Address Event Representation (AER) for implementation of the biCNN model. (A) Example of AER for a network with four neurons. (B) Vectors required for representing the network in (A) by using AER. The specific case for neuron #3 is shown in red. N stands for the number of neurons (N = 4 in the example), and Nsyn, the number of synapses in the network (Nsyn = 6 in the example).
Mentions: Address Event Representation (AER) is a communication technique for sparse networks and has been successfully extrapolated to neural networks (Johansson and Lansner, 2007). In AER, four vectors are required to describe the network architecture (Figure 4B). The first vector ID encodes the neurons in the network, assigning each a unique ID. The second vector NP stores the number of pre-synapses for each neuron, and the third and fourth vectors P and W encode the IDs of the pre-synaptic neurons and the corresponding synaptic weights in a stacked, ordered way. For example, in Figure 4A, neuron #3 is contacted by two neurons (#2, #1); this information is clearly observed in the third element of vectors ID[2] = 3, NP[2] = 2 (red arrows Figure 4B). By accumulating the number of synapses for the neurons that precede the neuron #3 (neuron #2, 2 pre-synapses, neuron #1, 1 pre-synapse), the index for the pre-synapses of neuron #3 can be read in the vectors P[3] = 1, and P[4] = 2, with their respective weights in W[3] = 0.3, and W[4] = −0.4.

Bottom Line: The bi-hemispheric structure is inspired by the observation that lesioning one hemisphere reduces motor performance asymmetrically.Our results showed that asymmetrical conditions were successfully handled by the biCNN model, in contrast to a single hemisphere model or a classical non-adaptive proportional and derivative controller.Thus, we conclude that a bi-hemispheric structure and adequate spontaneous activity of cf inputs are critical for cerebellar asymmetrical motor learning.

View Article: PubMed Central - PubMed

Affiliation: Neural Cybernetics Laboratory, Department of Computer Science, Chubu University Kasugai, Japan.

ABSTRACT
To acquire and maintain precise movement controls over a lifespan, changes in the physical and physiological characteristics of muscles must be compensated for adaptively. The cerebellum plays a crucial role in such adaptation. Changes in muscle characteristics are not always symmetrical. For example, it is unlikely that muscles that bend and straighten a joint will change to the same degree. Thus, different (i.e., asymmetrical) adaptation is required for bending and straightening motions. To date, little is known about the role of the cerebellum in asymmetrical adaptation. Here, we investigate the cerebellar mechanisms required for asymmetrical adaptation using a bi-hemispheric cerebellar neuronal network model (biCNN). The bi-hemispheric structure is inspired by the observation that lesioning one hemisphere reduces motor performance asymmetrically. The biCNN model was constructed to run in real-time and used to control an unstable two-wheeled balancing robot. The load of the robot and its environment were modified to create asymmetrical perturbations. Plasticity at parallel fiber-Purkinje cell synapses in the biCNN model was driven by error signal in the climbing fiber (cf) input. This cf input was configured to increase and decrease its firing rate from its spontaneous firing rate (approximately 1 Hz) with sensory errors in the preferred and non-preferred direction of each hemisphere, as demonstrated in the monkey cerebellum. Our results showed that asymmetrical conditions were successfully handled by the biCNN model, in contrast to a single hemisphere model or a classical non-adaptive proportional and derivative controller. Further, the spontaneous activity of the cf, while relatively small, was critical for balancing the contribution of each cerebellar hemisphere to the overall motor command sent to the robot. Eliminating the spontaneous activity compromised the asymmetrical learning capabilities of the biCNN model. Thus, we conclude that a bi-hemispheric structure and adequate spontaneous activity of cf inputs are critical for cerebellar asymmetrical motor learning.

Show MeSH
Related in: MedlinePlus