Efficient Hybrid Training Method for Neuromorphic Hardware Using Analog Nonvolatile Memory.

Area of Science:

• Neuromorphic engineering
• Artificial intelligence hardware

Background:

• Neuromorphic hardware with analog synaptic devices offers energy and time efficiency for vector-matrix multiplication (VMM).
• Existing training methods for this hardware suffer from reduced accuracy due to analog device nonidealities and high training costs associated with conductance tuning protocols.

Purpose of the Study:

• To propose and experimentally demonstrate a novel hybrid training method for neuromorphic hardware that overcomes the limitations of current training approaches.
• To reduce the cost and complexity of training nonvolatile analog synaptic devices while maintaining high accuracy.

Main Methods:

• Development of a hybrid training method that bypasses the need for conductance tuning protocols to update weights in analog synaptic devices.
• Experimental validation of the proposed training method using fabricated neuromorphic hardware.

Main Results:

• The hybrid training method significantly reduces online training costs.
• Hardware-based neural network accuracy closely approaches software-based accuracy after just one epoch of training, even when only the first synaptic layer is trained.
• The method demonstrates effectiveness with various synaptic devices exhibiting highly nonlinear weight update characteristics.

Conclusions:

• The proposed hybrid training method offers an efficient and cost-effective solution for training neuromorphic hardware.
• This approach enhances the viability of neuromorphic hardware utilizing nonvolatile analog memory cells as a promising platform for future artificial intelligence.
• The method's applicability to low-power neuromorphic systems and diverse synaptic devices highlights its broad potential.