A Multilayer-Learning Current-Mode Neuromorphic System With Analog-Error Compensation.

Area of Science:

• Neuromorphic engineering
• Energy-efficient hardware design
• Machine learning hardware acceleration

Background:

• The proliferation of Internet-of-Things (IoT) applications necessitates specialized, energy-efficient hardware for machine learning (ML) tasks.
• Existing systems often struggle to balance learning and inference capabilities with low power consumption at the endpoint.
• The demand for adaptive systems that can process data locally is rapidly increasing.

Purpose of the Study:

• To present a novel multilayer-learning neuromorphic system incorporating an analog-based multiplier-accumulator (MAC) for energy-efficient ML.
• To demonstrate the system's capability to perform both forward and backward data processing for learning.
• To validate the system's performance and energy efficiency without relying on analog calibration circuits.

Main Methods:

• Development of a current-mode MAC processor fabricated using 28-nm CMOS technology.
• Implementation of a crossbar structure with 500 × 500 6-b transposable SRAM arrays for the MAC processor.
• Verification of the two-layer neural network system using prototype chips and a Field-Programmable Gate Array (FPGA).
• Utilizing stochastic gradient descent for learning with 1-b batch updates on 6-b synaptic weights.

Main Results:

• The proposed neuromorphic system achieved a 96.6% recognition rate on the MNIST dataset.
• A peak energy efficiency of 2.99 TOPS/W was recorded (1 OP = one 8-b unsigned × 6-b signed MAC operation).
• The system successfully compensated for analog non-idealities through the learning process, eliminating the need for calibration circuits.

Conclusions:

• The developed multilayer-learning neuromorphic system offers a highly energy-efficient solution for edge AI and IoT applications.
• Analog-based MACs, when integrated into a learning system, can overcome non-idealities and achieve high performance.
• This approach paves the way for more powerful and adaptive ML capabilities directly on resource-constrained devices.