Recurrent models of orientation selectivity enable robust early-vision processing in mixed-signal neuromorphic...

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Mixed-signal analog/digital neuromorphic circuits offer real-time simulation of biological neural dynamics.
• These circuits, like biological neurons, exhibit limited resolution and significant variability.
• Existing models often struggle with heterogeneity and resource optimization.

Purpose of the Study:

• To investigate how noisy and heterogeneous neuromorphic circuits can form functional neural processing units.
• To model the retinocortical visual pathway using a recurrent spiking neural network.
• To demonstrate the generation of tuned receptive fields on a non-ideal hardware substrate.

Main Methods:

• Developed a recurrent spiking neural network (RSNN) model.
• Simulated the retinocortical visual pathway on a mixed-signal neuromorphic substrate.
• Analyzed the properties of receptive fields generated by the model.

Main Results:

• The model successfully produced linear receptive fields tuned to specific orientations and spatial frequencies.
• Generated Gabor-like receptive fields with varying phase symmetries, outperforming feed-forward schemes.
• Demonstrated optimal utilization of hardware resources (synaptic connections, neuron count).

Conclusions:

• Noisy and heterogeneous neuromorphic circuits can achieve robust sensory processing.
• Principles of neural computation enable effective visual processing despite hardware limitations.
• The approach validates the use of analog circuits and memristive devices for bio-realistic neural systems.