Neural network design for energy-autonomous artificial intelligence applications using temporal encoding.

Area of Science:

• Artificial Intelligence
• Computer Engineering
• Microelectronics

Background:

• Micro-edge AI applications like sensors and wearables require energy autonomy, often relying on power sources with inherent voltage variations.
• Existing neural network (NN) hardware struggles to maintain performance under unpredictable power supply fluctuations.
• Robust operation over a dynamic power envelope is crucial for reliable micro-edge AI.

Purpose of the Study:

• To propose a novel neural network design approach resilient to variable power supplies.
• To develop a pulse width modulation (PWM)-based perceptron as a fundamental building block for NNs.
• To demonstrate the power elasticity and efficiency of PWM-based NNs for AI applications.

Main Methods:

• Utilized the principle of pulse width modulation (PWM) for NN design, encoding information in duty cycle values.
• Developed a novel arithmetic realization method in the PWM domain for perceptron circuits.
• Analyzed a 3x3 perceptron circuit and scaled the approach to complex multi-layer NNs.

Main Results:

• Demonstrated a PWM-based perceptron capable of performing arithmetic operations independently of carrier signal voltage and frequency.
• Successfully implemented and analyzed multi-layer NNs using the PWM approach.
• Validated the power elasticity, resilience, and efficiency of the proposed NN design using handwritten character recognition under significant voltage variations.

Conclusions:

• The proposed PWM-based NN design offers a robust solution for energy-autonomous AI at the micro-edge.
• This approach enables reliable NN hardware operation despite dynamic power envelope challenges.
• PWM-based NNs present a promising direction for harmonizing energy-autonomous computing and artificial intelligence.