A spiking neural network for active efficient coding

Area of Science:

• Computational Neuroscience
• Robotics
• Artificial Intelligence

Background:

• Biological vision systems learn to encode visual input and control eye movements simultaneously.
• The Active Efficient Coding (AEC) framework models this joint learning using traditional cameras.
• Event-based cameras offer advantages over frame-based systems due to their retina-inspired design.

Purpose of the Study:

• To propose and implement the first Active Efficient Coding (AEC) system using Spiking Neural Networks (SNNs) driven by event-based camera input.
• To demonstrate efficient visual representation learning and motor command generation within a fully spiking neural network framework.
• To explore the capabilities of this novel system in tasks requiring dynamic visual interaction.

Main Methods:

• Developed a two-layer Spiking Neural Network (SNN) for efficient encoding of event-based camera data.
• Integrated this SNN with a spiking reinforcement learner to generate motor commands.
• Designed an intrinsic reward signal computed from the SNN's activity levels to guide learning.
• Evaluated the system on visual tracking and orientation stabilization tasks.

Main Results:

• Successfully implemented a fully Spiking Neural Network (SNN) based Active Efficient Coding (AEC) system.
• Demonstrated the system's ability to perform visual tracking of translating targets.
• Showcased the system's capability in stabilizing the orientation of rotating targets.
• Achieved joint learning of visual representations and motor control within the SNN framework.

Conclusions:

• This work presents the first fully spiking Active Efficient Coding (AEC) model.
• The proposed system effectively utilizes event-based cameras and Spiking Neural Networks (SNNs) for autonomous visual learning and control.
• This approach holds promise for developing more efficient and biologically plausible artificial vision systems.