Neuroinspired unsupervised learning and pruning with subquantum CBRAM arrays.

Area of Science:

• Neuromorphic Engineering
• Materials Science

Background:

• Resistive RAM (ReRAM) crossbar arrays reduce data transfer and enable parallel computation for neural networks.
• Energy efficiency is critical for deploying neural networks in power-limited environments.

Purpose of the Study:

• To develop a hardware/software co-design approach for energy-efficient neural network computation using low-energy subquantum conductive bridging RAM (CBRAM) devices.
• To implement a network pruning technique that reduces energy consumption at the network level during training.

Main Methods:

• Demonstration of low-energy subquantum CBRAM devices with gradual switching characteristics suitable for hardware weight updates.
• Development of a network pruning algorithm applicable during the training phase.
• Experimental validation using a 512 kbit subquantum CBRAM array for unsupervised learning on the MNIST dataset.

Main Results:

• Subquantum CBRAM devices exhibit gradual switching, crucial for unsupervised learning weight updates.
• The developed network pruning algorithm effectively reduces energy consumption during training.
• High recognition accuracy was achieved on the MNIST dataset using the digital implementation of unsupervised learning on the CBRAM array.

Conclusions:

• The hardware/software co-design approach, integrating subquantum CBRAM and network pruning, significantly enhances energy efficiency for neural networks.
• This work paves the way for resistive memory-based neuro-inspired systems capable of autonomous learning in power-constrained settings.