Task-structured modularity emerges in artificial networks and aligns with brain architecture

Problem

This work addresses a gap in understanding how the functional demands of learning complex tasks influence the topology of artificial neural networks. The authors explore the emergence of task-structured modularity in networks, which has implications for both artificial intelligence and neuroscience. The paper is a preprint and has not undergone peer review.

Method

The authors utilize Recurrent Neural Networks (RNNs) as their primary architecture. They investigate two training paradigms: multitask learning and single-task training. The data is sourced from the Human Connectome Project (HCP), encompassing 84 cortical areas. While the specific training compute and loss function are not disclosed, the study emphasizes that incremental multitask learning yields the highest degree of modularity and superior performance compared to single-task training.

Results

The results indicate that multitask learning significantly enhances network modularity when compared to single-task training. Additionally, the structural properties of the task-induced networks exhibit a closer resemblance to biological brain networks than to models constrained by spatial parameters, specifically when contrasted with spatial constraint models.

Limitations

The authors do not report any limitations in their study. However, the lack of specified training compute and loss function may hinder reproducibility and further analysis.

Why it matters

The findings have significant implications for the design of artificial neural networks, suggesting that incorporating multitask learning can lead to more biologically plausible architectures. This work may inform future research in both AI and neuroscience, particularly in understanding how complex task learning can shape network structures.