Vectorized instructive signals in cortical dendrites during a brain-computer interface task

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Vectorized teaching signals are crucial for modern machine learning algorithms like backpropagation.
• Theoretical models propose that neural circuits might implement vectorized learning at the cellular level using dendritic compartments.
• The biological basis for solving the credit assignment problem in the brain via vectorized signals remains largely untested.

Purpose of the Study:

• To investigate the existence and function of vectorized instructive signals within cortical dendrites.
• To test the hypothesis that separate dendritic compartments process feedforward and feedback information for learning.
• To explore how neural circuits in the brain might implement credit assignment.

Main Methods:

• Utilized a neurofeedback brain-computer interface (BCI) task with mice.
• Trained mice to modulate activity of retrosplenial cortex layer 5 pyramidal neurons to control a visual grating.
• Recorded neural activity from neuronal somas and distal apical dendrites using GCaMP imaging.

Main Results:

• Somatic and dendritic signal magnitudes contained information about task variables like reward and error, acting as instructive signals.
• The nature of these teaching signals depended on individual neurons' causal role in the task and predicted learning-related activity changes.
• Optogenetic disruption of these dendritic signals impaired learning, providing causal evidence.

Conclusions:

• Presented the first biological evidence for vectorized instructive signals in the brain, implemented in cortical dendrites.
• Demonstrated that dendrites perform semi-independent computations, supporting a mechanism for solving credit assignment.
• Unveiled a potential cellular-level solution for credit assignment in biological neural networks.