In situ Parallel Training of Analog Neural Network Using Electrochemical Random-Access Memory

Area of Science:

• Materials Science
• Computer Science
• Electrical Engineering

Background:

• In-memory computing with non-volatile resistive memory offers energy efficiency for artificial neural networks.
• Accurate in situ training is hindered by nonlinear and stochastic switching in resistive memory elements.
• Electrochemical random-access memory (ECRAM), or redox transistors, show promise due to low write currents and linear switching.

Purpose of the Study:

• To experimentally demonstrate accurate in situ parallel training of an ECRAM-based neural network.
• To validate simulation predictions of ECRAM's potential for energy-efficient artificial neural network training.
• To advance the development of ECRAM for large-scale artificial neural network accelerators.

Main Methods:

• A 3x3 array of ECRAM devices was trained in situ to discriminate logic gates (AND, OR, NAND).
• Outer product updates were used for parallel, on-line training, recording synaptic weight evolution.
• Device switching characteristics were analyzed for linearity and reproducibility.

Main Results:

• The ECRAM array successfully learned to discriminate logic gates through parallel in situ training.
• Simulations accurately predicted convergence epochs and quantitatively matched individual device weight evolution.
• Linear and reproducible switching of ECRAM devices was confirmed during training.

Conclusions:

• This work presents the first experimental demonstration of in situ parallel training in an ECRAM crossbar array.
• The strong agreement between experimental results and simulations validates ECRAM's suitability for neural network acceleration.
• ECRAM technology is a significant step toward highly energy-efficient deep neural networks.