Disentangling Superpositions: Interpretable Brain Encoding Model with Sparse Concept Atoms

Area of Science:

• Neuroscience
• Cognitive Science
• Artificial Intelligence

Background:

• Encoding models using artificial neural network (ANN) features predict brain responses to stimuli.
• Interpreting these models is challenging due to feature superposition and entanglement in dense embeddings.
• This entanglement prevents clear identification of semantic features and voxel selectivity.

Purpose of the Study:

• To develop a novel encoding model for enhanced interpretability of brain responses.
• To address the limitation of feature entanglement in dense embeddings.
• To enable direct readout of conceptual selectivity from voxel weights.

Main Methods:

• Introduced the Sparse Concept Encoding Model (SCEM).
• Transformed dense embeddings into a higher-dimensional, sparse, non-negative space of learned concept atoms.
• Applied the SCEM to functional magnetic resonance imaging (fMRI) data from story listening.

Main Results:

• The SCEM achieved prediction performance comparable to conventional dense models.
• The model significantly enhanced the interpretability of neural representations.
• Enabled disentanglement of overlapping cortical representations (e.g., time, space, number).

Conclusions:

• The SCEM provides a scalable and interpretable bridge between ANN features and brain representations.
• Offers a framework for novel neuroscientific analyses of conceptual maps.
• Facilitates a deeper understanding of semantic feature encoding in the human brain.