Sign backpropagation: An on-chip learning algorithm for analog RRAM neuromorphic computing systems.

Area of Science:

• Neuromorphic Computing
• Energy-Efficient Hardware

Background:

• Deep learning models demand substantial computational resources and energy.
• Resistive random access memory (RRAM) presents a promising avenue for scalable, low-power neural networks.
• Variability in RRAM devices complicates the direct porting of high-precision neural networks from digital systems.

Purpose of the Study:

• To develop an on-chip learning algorithm for RRAM-based neural networks.
• To address the challenge of integrating digital periphery computations with analog RRAM crossbars.
• To enable high-precision performance in RRAM-based multilayer perceptrons (MLPs).

Main Methods:

• Proposed a novel on-chip learning algorithm: sign backpropagation (SBP).
• Implemented SBP for RRAM-based MLPs utilizing binary interfaces (0, 1) for forward and 2-bit (±1, 0) for backward processes.
• Simulated the proposed method and architecture on the MNIST dataset.

Main Results:

• Achieved comparable classification accuracy to conventional MLPs on the MNIST dataset.
• Demonstrated significant savings in area and energy consumption by optimizing intermediate result storage.
• Leveraged the inherent advantages of RRAM crossbars for efficient neuromorphic computation.

Conclusions:

• The sign backpropagation algorithm is effective for RRAM-based neural networks.
• The proposed architecture offers a viable solution for energy-efficient and scalable deep learning.
• This approach effectively bridges the gap between digital and analog computing for AI hardware.