Generalized reconfigurable memristive dynamical system (MDS) for neuromorphic applications

Area of Science:

• Neuromorphic engineering
• Computational neuroscience
• Circuit design

Background:

• Neuromorphic dynamical systems require efficient hardware implementations.
• Existing Cellular Memristive Dynamical Systems (CMDS) often involve numerous switches and memristors.
• Flexible and programmable hardware is crucial for diverse neuromorphic applications.

Purpose of the Study:

• To present a novel general cellular mapping scheme for 2D neuromorphic dynamical systems.
• To develop an efficient mixed analog-digital circuit for hardware implementation of the scheme.
• To demonstrate the circuit's capability in simulating various neuron models and its potential for learning systems.

Main Methods:

• Developed a novel cellular mapping scheme for neuromorphic systems.
• Designed a hybrid memristor-crossbar/CMOS circuit using 4n memristors and no switches per n-cell.
• Simulated FitzHugh-Nagumo (FHN), Adaptive Exponential (AdEx), and Izhikevich neuron models.
• Conducted dynamical response and error analyses, and built a hardware prototype.

Main Results:

• The proposed scheme uses fewer components (4n memristors, no switches) compared to CMDS (2n memristors, 2n switches).
• The circuit supports both analog and one-hot digital dynamical variables, enabling versatile networking.
• Simulations of FHN, AdEx, and Izhikevich models showed accurate dynamical behaviors based on bifurcation scenarios.
• Error analysis confirmed the approach's suitability and accuracy.

Conclusions:

• The novel scheme and circuit offer an efficient, programmable, and accurate hardware platform for neuromorphic dynamical systems.
• The approach is suitable for implementing bio-inspired neuron models and can be applied to learning systems and analytically indescribable dynamics.
• The developed hardware prototype validates the feasibility and effectiveness of the proposed approach.