Predictive Coding Light

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Current machine learning systems are highly energy-intensive compared to biological brains.
• Neuromorphic systems aim to reduce energy consumption by mimicking the brain's spiking neuron information coding.
• Effective learning strategies for energy-efficient processing in spiking neural networks remain an open challenge.

Purpose of the Study:

• To propose Predictive Coding Light (PCL), a novel recurrent hierarchical spiking neural network.
• To enable unsupervised representation learning with enhanced energy efficiency.
• To investigate biologically plausible learning rules for spiking neural networks.

Main Methods:

• Developed Predictive Coding Light (PCL), a recurrent hierarchical spiking neural network.
• Implemented a novel predictive coding approach that suppresses predictable spikes instead of transmitting prediction errors.
• Utilized biologically plausible spike-timing based learning rules.

Main Results:

• PCL achieves unsupervised representation learning by transmitting a compressed input representation.
• The model successfully reproduces findings on information processing in the visual cortex.
• PCL demonstrates strong performance in downstream classification tasks.

Conclusions:

• PCL offers a new, energy-efficient approach to predictive coding in spiking neural networks.
• The proposed method is implementable in both artificial and natural neural systems.
• This work advances the understanding of efficient information processing in neural networks.