Photons guided by axons may enable backpropagation-based learning in the brain

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biophysics

Background:

• Understanding brain learning algorithms remains incomplete despite advances in synaptic plasticity and neuron function.
• Artificial neural networks utilize backpropagation for learning, but its biological plausibility in the brain is debated due to extensive feedback requirements.

Purpose of the Study:

• To propose biophotons as a biological mechanism for backward information transmission in neural networks.
• To investigate the potential of biophotons in realizing a biologically plausible backpropagation-like learning algorithm.

Main Methods:

• A novel algorithm using stochastic photonic feedback for backward transmission of teaching signals was developed.
• A three-layered artificial neural network was trained on the MNIST handwritten digit classification task.
• The model incorporated realistic constraints, including low biophoton emission rates, limited information per photon, and noise.

Main Results:

• The proposed system successfully learned the MNIST classification task, demonstrating the efficacy of biophoton-mediated backward transmission.
• Learning was achieved even with low biophoton emission rates and information-limited (one bit per photon) feedback.
• The model showed robustness in the presence of noise photons, suggesting a viable biological learning mechanism.

Conclusions:

• Biophotons may serve a functional role in backward information transmission within the brain.
• This research presents a novel, biologically plausible alternative to traditional backpropagation for neural learning.
• The findings open new avenues for understanding brain computation and developing advanced artificial intelligence.