A CMOS feedforward neural-network chip with on-chip parallel learning for oscillation cancellation.

Area of Science:

• Mixed-signal integrated circuit design
• Artificial intelligence hardware
• Control systems engineering

Background:

• Real-time adaptation in control systems is crucial for dynamic environments.
• Existing neural network chips often require complex error calculation methods.
• On-chip learning algorithms are needed for efficient, autonomous system control.

Purpose of the Study:

• To present a mixed-signal CMOS feedforward neural-network chip with on-chip error reduction.
• To implement and evaluate a genetic random search algorithm (Random Weight Change - RWC) for high-speed parallel learning.
• To demonstrate the chip's capability as a direct feedback controller for real-time instability suppression.

Main Methods:

• Designed and fabricated a mixed-signal CMOS feedforward neural-network chip.
• Integrated on-chip error-reduction hardware for real-time adaptation.
• Implemented the Random Weight Change (RWC) algorithm, a genetic random search method, for on-chip learning.
• Conducted hardware experiments using the chip as a direct feedback controller.

Main Results:

• The chip features compact on-chip weights enabling high-speed parallel learning.
• The RWC algorithm facilitates learning without requiring a known desired output, suitable for direct feedback control.
• Hardware experiments successfully demonstrated the RWC chip's ability to suppress unstable oscillations in real time.
• The chip effectively modeled and controlled combustion engine instability.

Conclusions:

• The developed mixed-signal CMOS neural-network chip with RWC algorithm offers efficient real-time adaptation.
• The chip serves as an effective direct feedback controller, capable of suppressing complex system instabilities.
• This work highlights a promising approach for autonomous control systems in dynamic environments.