Wavelet-like receptive fields emerges by non-linear minimization of neuron error

Area of Science:

• Computational Neuroscience
• Machine Learning
• Signal Processing

Background:

• Redundancy reduction in neural coding has been a significant research area since the 1960s.
• Current prominent strategies focus on mutually independent output signals.
• A gap exists in efficiently handling non-orthogonal (dependent) signals.

Purpose of the Study:

• To propose a novel strategy for redundancy reduction that accommodates non-orthogonal signals.
• To develop a computationally efficient neural network algorithm.
• To provide a biologically plausible framework for single neuron redundancy reduction.

Main Methods:

• Designed a neuron model employing non-linearity on the error, rather than the usual squared error.
• Developed an algorithm capable of processing dependent signals.
• Validated the algorithm's biological plausibility and emergent properties.

Main Results:

• The proposed method effectively handles non-orthogonal signals, overcoming limitations of existing independent component analysis approaches.
• The error non-linearity approach proved computationally more economical and avoided permutation/scaling issues.
• Wavelet-like receptive fields emerged from natural images processed by the algorithm, suggesting biological relevance.

Conclusions:

• The developed algorithm offers an effective and efficient method for redundancy reduction, particularly for dependent signals.
• The approach aligns with the single neuron redundancy reduction doctrine and exhibits biologically plausible emergent properties.
• This work advances neural coding strategies with potential applications in both neuroscience and machine learning.