A High-Accuracy and Energy-Efficient CORDIC Based Izhikevich Neuron With Error Suppression and Compensation

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Digital Hardware Design

Background:

• Bio-inspired neuron models are crucial for brain-like neural networks in neuroscience and neuromorphic applications.
• Efficient hardware design of these models faces challenges in balancing accuracy, energy consumption, and cost.
• Existing designs often struggle to optimize these critical parameters simultaneously.

Purpose of the Study:

• To propose a novel, high-accuracy, and energy-efficient Izhikevich neuron design using Fast-Convergence COordinate Rotation DIgital Computer (FC-CORDIC).
• To systematically analyze and reduce errors in the Izhikevich neuron model for improved hardware implementation.
• To enhance the performance of neuromorphic hardware by optimizing neuron model efficiency.

Main Methods:

• Developed an error propagation model for systematic analysis and reduction of Izhikevich neuron design errors.
• Introduced Parameter-Tuning Error Compensation (PTEC) and Bitwidth-Extension Error Suppression (BEES) methods for error mitigation.
• Utilized FC-CORDIC for square calculations, removing redundant iterations and reducing accumulated errors and computation.
• Proposed an optimized fixed-point FC-CORDIC design to minimize hardware overhead while maintaining accuracy.

Main Results:

• The FC-CORDIC based Izhikevich neuron design demonstrates significant improvements in accuracy and energy efficiency.
• Error reduction techniques (PTEC and BEES) effectively minimize inaccuracies in the neuron model.
• The use of FC-CORDIC reduces computational complexity and accumulated errors compared to conventional methods.
• Field-Programmable Gate Array (FPGA) implementation confirms high performance with acceptable hardware overhead.

Conclusions:

• The proposed FC-CORDIC based Izhikevich neuron design offers a superior balance of accuracy, energy efficiency, and hardware cost.
• This design represents a significant advancement for implementing efficient and accurate brain-like neural networks.
• The methods presented contribute to the state-of-the-art in neuromorphic engineering and computational neuroscience hardware.