An Approximation of the Error Backpropagation Algorithm in a Predictive Coding Network with Local Hebbian Synaptic...

Area of Science:

• Computational neuroscience
• Machine learning

Background:

• Cortical networks require synaptic weight updates across hierarchical levels for efficient learning.
• Error backpropagation is effective but biologically implausible due to its reliance on non-local information.
• Existing biologically plausible models often need complex control or plasticity rules.

Purpose of the Study:

• To demonstrate that predictive coding networks can perform autonomous supervised learning.
• To show that simple, local Hebbian plasticity can be sufficient for this learning.
• To investigate the convergence of predictive coding weight updates to backpropagation.

Main Methods:

• Developing a network within the predictive coding framework.
• Employing simple local Hebbian plasticity for synaptic updates.
• Analyzing the autonomous learning capabilities and weight change convergence.

Main Results:

• The predictive coding network achieved efficient supervised learning autonomously.
• The learning process utilized only simple local Hebbian plasticity.
• Under specific parameters, weight changes converged towards those of the backpropagation algorithm.

Conclusions:

• Predictive coding networks offer a biologically plausible mechanism for efficient supervised learning.
• Simple Hebbian plasticity may be sufficient for hierarchical synaptic modifications.
• This framework supports the possibility of cortical networks implementing efficient learning through local rules.