Modular representations emerge in neural networks trained to perform context-dependent tasks

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Cognitive science

Background:

• The brain exhibits large-scale modularity in regions, influenced by connectivity and physical geometry.
• Debate exists on whether specialized neuronal sub-populations (modules) exist within single brain regions.

Purpose of the Study:

• To investigate the emergence and function of local modularity in artificial neural networks.
• To determine conditions under which modularity supports context-dependent behavior and learning.

Main Methods:

• Studied artificial neural networks with varying input dimensionality.
• Analyzed the emergence of modular specialization at the population level.
• Examined the relationship between modularity and representational properties.

Main Results:

• Local modularity emerges in neural networks supporting context-dependent behavior, specifically with low-dimensional input, without anatomical constraints.
• Modular specialization leads to abstract representations, enabling rapid learning and generalization on novel and related tasks.
• Non-modular representations facilitate faster learning of unrelated contexts.

Conclusions:

• Local neural network modularity is driven by input dimensionality and supports adaptive behavior.
• Findings reconcile conflicting experimental data on neuronal specialization.
• Predicts future experimental avenues for investigating brain modularity.