Symmetric Predictive Estimator for Biologically Plausible Neural Learning.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Neural Networks

Background:

• Perception involves bidirectional neural processing, utilizing both feedforward and feedback pathways.
• Existing neural network learning rules are often limited to feedforward networks or rely on biologically implausible assumptions like weight copying.
• Predictive estimators (PEs) offer a potential solution for bidirectional learning but still depend on weight copying.

Purpose of the Study:

• To develop a biologically plausible learning rule for neural networks that mimics bidirectional brain processes.
• To address the limitations of current neural network learning methods, particularly their inability to handle feedback pathways effectively without unrealistic assumptions.
• To propose a novel architecture, the symmetric PE (SPE), capable of learning both feedforward and feedback weights using only local information.

Main Methods:

• Introduction of the symmetric predictive estimator (SPE) architecture.
• Demonstration of the SPE's ability to learn feedforward and feedback connection weights independently.
• Utilizing only locally available information for weight updates, eliminating the need for weight copying.

Main Results:

• The SPE successfully learns complex mappings without employing weight copying.
• SPE networks exhibit the capacity for bidirectional information processing, mirroring real brain function.
• The proposed architecture shows potential for application in deeper neural network models.

Conclusions:

• The symmetric predictive estimator (SPE) provides a biologically plausible mechanism for learning bidirectional neural computations.
• SPE overcomes the limitations of previous methods by learning feedforward and feedback pathways without weight copying.
• SPE networks represent a promising advancement for developing more realistic and capable artificial neural systems.