Threshold learning algorithm for memristive neural network with binary switching behavior.

Area of Science:

• Neuromorphic Computing
• Materials Science

Background:

• On-chip learning in neuromorphic systems faces challenges from hardware nonidealities like nonlinearity and limited weight adjustment.
• Memristor crossbar arrays offer potential for efficient neural network implementation but require robust learning algorithms.

Purpose of the Study:

• To propose and experimentally validate a variation-tolerant threshold learning algorithm for ternary neural networks utilizing memristor crossbar arrays.
• To address limitations in fine weight adjustment and hardware nonidealities in on-chip learning.

Main Methods:

• Developed a threshold learning algorithm using two distinct memristor resistance states (HRS and LRS) for weight representation.
• Programmed a 32x32 memristor crossbar array, achieving tightly separated resistance states.
• Trained a 64x10 single-layer fully connected network on an 8x8 MNIST dataset using the proposed algorithm.

Main Results:

• Successfully programmed distinct high-resistance (HRS) and low-resistance (LRS) states in the memristor crossbar array.
• Demonstrated a trained network with only a 0.42% drop in classification accuracy compared to baseline results.
• The threshold learning algorithm effectively updated weights when backpropagation gradients exceeded a set threshold.

Conclusions:

• The proposed threshold learning algorithm is effective for variation-tolerant ternary neural networks in memristor crossbar arrays.
• This approach alleviates programming burdens and is suitable for future neuromorphic architectures.
• Experimental validation shows high performance with minimal accuracy degradation, highlighting the algorithm's practical applicability.