A memristive synaptic circuit and optimization algorithm for synaptic control

Area of Science:

• Materials Science
• Computer Science
• Electrical Engineering

Background:

• Backpropagation training in machine learning requires linear weight changes for direct application to memristor crossbar arrays.
• Non-linear memristance and its temporal instability pose memory and energy challenges for direct training.
• Existing methods require complex algorithms or memory to handle memristor non-linearity.

Purpose of the Study:

• To develop a method for achieving linear weight control in memristor circuits for machine learning.
• To overcome the memory and energy drawbacks associated with non-linear memristance.
• To enable direct application of backpropagation training to memristor crossbar arrays.

Main Methods:

• Utilized an optimization method with charge-controlled and flux-controlled memristor equations.
• Employed the artificial bee colony algorithm to determine circuit parameters and control signal duration.
• Focused on achieving linear control of weight changes for positive and negative weights.

Main Results:

• Achieved linear control of weight change with a sensitivity up to 0.02.
• Experimental weight control demonstrated a mean square error of 2.33 x 10^-4.
• Attained a software-based test accuracy tracking rate of 98.186%.

Conclusions:

• The proposed optimization method and cost function enable linear control for online training with memristor elements.
• This approach simplifies weight control, addressing non-linearity issues in memristor-based machine learning.
• The findings pave the way for more efficient and direct application of machine learning training on memristor hardware.