Random synaptic feedback weights support error backpropagation for deep learning.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Neuroscience

Background:

• The brain's layered neural processing is powerful but learning is complex due to difficulty in identifying error sources.
• Machine learning uses backpropagation for error assignment, requiring precise symmetric backward connectivity, which is biologically implausible in the brain.

Purpose of the Study:

• To investigate if effective error propagation in deep neural networks can occur without the strict architectural constraints of backpropagation.
• To propose and validate a simpler mechanism for transmitting teaching signals across multiple neuronal layers in the brain.

Main Methods:

• Developed a novel error propagation mechanism that utilizes random synaptic weights instead of precise symmetric connectivity.
• Tested the performance of this new mechanism against traditional backpropagation across various computational tasks.

Main Results:

• The proposed mechanism effectively assigns blame and transmits teaching signals across multiple layers, performing comparably to backpropagation.
• Demonstrated that precise, symmetric backward connectivity is not a necessary constraint for effective error propagation in learning.

Conclusions:

• The findings challenge long-held assumptions about the algorithmic constraints on neural learning in the brain.
• This work opens new avenues for understanding how biological neural networks might implement error-driven learning signals.