A cerebellum inspired spiking neural network as a multi-model for pattern classification and robotic trajectory...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Robotics

Background:

• Spiking neural networks (SNNs) process spatiotemporal information, finding applications in pattern recognition.
• Cerebellar architecture inspires bio-inspired networks for event-driven tasks, particularly in motor learning.
• Deep learning networks (DLNs) offer learning rules applicable to neural network abstractions.

Purpose of the Study:

• To implement a cerebellum-inspired SNN model incorporating cerebellar neuron dynamics and learning mechanisms.
• To evaluate the model's pattern discrimination and trajectory optimization capabilities on ML datasets and robotic control tasks.
• To assess the generalization performance and efficiency of the proposed SNN model compared to standard ML algorithms.

Main Methods:

• Developed a cerebellum-inspired SNN model with granular layer, Purkinje cell layer, and cerebellar nuclei dynamics.
• Tested pattern discrimination on standard ML datasets and trajectory following on a robotic articulator.
• Tuned the SNN for supervised learning to evaluate its classification and reconstruction accuracy.

Main Results:

• The SNN model achieved 72% accuracy in pattern classification, comparable to ML algorithms like MLP (78%) and libSVM-linear (85.7%).
• The model demonstrated generalized pattern classification, outperforming data-specific models on smaller datasets.
• Successfully reconstructed robotic arm trajectories with a low error rate (±3 cm) for a 6-DOF arm.

Conclusions:

• The cerebellum-inspired SNN offers generalized processing and efficient computation, reducing storage and increasing speed.
• The model shows potential for sensor-free robotic controllers and generalized pattern classification algorithms.
• This work contributes to motor learning theory and highlights the efficacy of bio-inspired SNNs.