Learning to combine top-down context and feed-forward representations under ambiguity with apical and basal dendrites

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Neurobiology

Background:

• Neocortex features top-down projections to sensory areas, hypothesized to provide context for resolving sensory ambiguity.
• Pyramidal neurons integrate inputs via distinct apical and basal dendrites, but computational models for this remain underexplored.

Purpose of the Study:

• To computationally demonstrate how distinct neuronal compartments can enable flexible contextual integration and sensory processing.
• To introduce a novel deep neural network architecture mimicking apical and basal dendrite functions.

Main Methods:

• Implemented a deep neural network with separate apical and basal compartments.
• Integrated top-down contextual information into apical compartments and bottom-up sensory data into basal compartments.
• Developed a new contextual integration task using generative modeling.

Main Results:

• Deep neural networks with the "apical prior" architecture outperformed single-compartment networks.
• A sparse subset of neurons received significant top-down contextual signals.
• Sparse gain modulation was identified as a necessary mechanism for performance.

Conclusions:

• The proposed "apical prior" architecture provides a viable computational framework for integrating contextual and sensory information.
• This model suggests a potential neural mechanism for handling real-world sensory ambiguities in animals.