CMOS realization of a 2-layer CNN universal machine chip.

Area of Science:

• Neuroscience and Computational Engineering
• Biologically inspired computing

Background:

• The biological retina exhibits complex spatiotemporal dynamics crucial for image processing.
• Cellular Neural Networks (CNNs) offer a computational model for mimicking retinal functions.
• Locally connected elementary nonlinear processors form the basis of these CNN models.

Purpose of the Study:

• To design and fabricate a prototype chip emulating biological retina features.
• To explore complex spatiotemporal dynamics for image processing applications.
• To present design challenges, trade-offs, and testing results of the fabricated chip.

Main Methods:

• Modeling biological retina features using a 2-layer cellular neural network (CNN).
• Designing and fabricating a prototype chip using 0.5 micrometer CMOS technology.
• Utilizing locally connected elementary nonlinear processors within the CNN architecture.
• Employing analog mode operation for a significant portion of the 0.5 x 10^6 transistors.

Main Results:

• Successful design and fabrication of a prototype CNN chip.
• Demonstration of the chip's capability to explore complex spatiotemporal dynamics.
• Analysis of design challenges and trade-offs inherent in the system.

Conclusions:

• The prototype chip successfully models aspects of the biological retina.
• The fabricated CNN chip provides a platform for investigating spatiotemporal dynamics in image processing.
• The study highlights the feasibility of analog CMOS technology for bio-inspired neural network hardware.