Blooming and pruning: learning from mistakes with memristive synapses.

Area of Science:

• Neuroscience
• Materials Science
• Computer Engineering

Background:

• Brain development involves synaptic pruning for environmental adaptation and cognitive skill formation.
• A 1999 learning scheme proposed error-based synaptic connection elimination.
• CMOS integrated HfO2-based memristive devices offer potential for neuromorphic computing.

Purpose of the Study:

• To implement a learning scheme inspired by brain's blooming and pruning in a neuromorphic circuit.
• To utilize HfO2-based memristive devices for hardware-based, local, and energy-efficient learning.
• To identify system and device parameters for a robust memristive neuromorphic circuit capable of association tasks.

Main Methods:

• Implementation of a two-layer neural network using CMOS integrated HfO2-based memristive devices.
• Development of a self-organized learning scheme without positive reinforcement, leveraging device variability.
• Combined experimental and simulation-based parameter study.

Main Results:

• Demonstration of a compact and robust memristive neuromorphic circuit.
• Successful implementation of hardware, local, and energy-efficient learning.
• Identification of key parameters for effective association task handling.

Conclusions:

• The HfO2-based memristive neuromorphic circuit effectively implements error-based learning.
• This approach offers a promising pathway for energy-efficient, self-organized learning hardware.
• The study provides a framework for designing robust memristive circuits for cognitive tasks.