Pulse-stream VLSI neural networks mixing analog and digital techniques

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Integrated Circuit Design

Background:

• The pulse-stream technique represents neural states as pulse sequences.
• Existing methods for pulsed encoding, arithmetic, and communication present challenges.
• Novel synapse designs are needed for efficient analog computation in neural networks.

Purpose of the Study:

• To review the pulse-stream technique and its associated challenges.
• To present and compare two novel synapse designs for analog neural computation.
• To demonstrate the feasibility of pulse-stream devices in practical applications.

Main Methods:

• Review of general issues in pulsed encoding, arithmetic, and intercommunication.
• Design and comparison of two synapse architectures: fully analog and pulse-stretching.
• Implementation of integrated circuits with up to 15,000 programmable synaptic connections.
• Development of a demonstrator project for robot localization using asynchronous, analog, pulse-stream devices.

Main Results:

• Two distinct synapse designs were developed and analyzed.
• The first design utilizes a fully analog computational approach with digital pulse signaling.
• The second design incorporates digital memory for weight storage and pulse-stretching for synapses.
• Integrated circuits with a high density of analog synaptic connections were fabricated.
• A robot localization network was successfully implemented using these pulse-stream devices.

Conclusions:

• The pulse-stream technique offers a viable approach for neuromorphic computing.
• The presented synapse designs provide efficient methods for analog computation.
• The successful demonstrator project highlights the potential of these devices for real-world applications.
• Further development in analog, pulse-stream hardware can advance the field of artificial intelligence.