Online dynamical learning and sequence memory with neuromorphic nanowire networks.

Area of Science:

• Neuromorphic computing
• Materials science
• Artificial intelligence

Background:

• Nanowire Networks (NWNs) are emerging neuromorphic systems utilizing nanostructured materials.
• NWNs exhibit resistive memory switching at junctions, mimicking synaptic plasticity.
• Previous research harnessed NWN neuromorphic dynamics for temporal learning.

Purpose of the Study:

• To demonstrate online learning from spatiotemporal dynamical features using NWNs.
• To apply NWNs to image classification and sequence memory recall tasks.
• To elucidate the role of memory in enhancing NWN learning.

Main Methods:

• Implementation of an NWN device for online learning experiments.
• Utilizing spatiotemporal dynamical features for learning tasks.
• Applying the device to the MNIST handwritten digit classification and sequence memory recall tasks.

Main Results:

• Achieved 93.4% accuracy on the MNIST handwritten digit classification task using online dynamical learning.
• Observed a correlation between classification accuracy and mutual information for individual digit classes.
• Demonstrated online learning and recall of spatiotemporal sequences through memory patterns in dynamical features.

Conclusions:

• Provided proof-of-concept for online learning from spatiotemporal dynamics in NWNs.
• Showcased the potential of NWNs for complex cognitive tasks like image classification and memory.
• Highlighted the significance of memory in enhancing the learning capabilities of neuromorphic systems.