Emergent Language Symbolic Autoencoder (ELSA) with weak supervision to model hierarchical brain networks

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Brain networks exhibit complex hierarchical organization.
• Current deep learning models often lack interpretability and struggle with hierarchical data.
• Representing brain network complexity is crucial for understanding brain function.

Purpose of the Study:

• To develop a novel hierarchical symbolic autoencoder, the Emergent Language Symbolic Autoencoder (ELSA).
• To represent brain networks as interpretable symbolic sentences using a hierarchical structure.
• To improve the interpretability and hierarchical modeling of brain network data.

Main Methods:

• Proposed the Emergent Language Symbolic Autoencoder (ELSA) architecture.
• Developed novel hierarchically-aware loss functions (Progressive, Strict, Containing Bias).
• Applied ELSA to resting-state fMRI data from the 1000 Functional Connectomes Project, using weak supervision from Independent Component Analysis (ICA) order.

Main Results:

• ELSA generated symbolic sentences with clear hierarchical organization, reflecting coarse-to-fine brain network structures.
• The Progressive Strict loss function significantly improved hierarchical consistency compared to baseline models.
• Achieved near-perfect hierarchical consistency at higher ICA orders and 43.5% at the lowest order, with qualitatively superior network reconstructions.

Conclusions:

• ELSA provides a transparent, multi-level description of functional brain organization by translating opaque feature vectors into an interpretable symbolic language.
• The framework offers a generalizable approach for studying other hierarchically structured biomedical data.
• ELSA enhances the interpretability of deep learning models for complex biological systems.