Summary: Current AI models largely imitate the cortex—the brain’s high-level, outer layer—but often neglect the older, deeper subcortical structures that mediate rapid reactions, emotion, and basic learning. A research team from the Netherlands proposes a new computational architecture that integrates these subcortical pathways with cortical hierarchies to produce more flexible, efficient, and biologically plausible AI systems.
The study shows that adding a fast, shallow subcortical route alongside the deep, hierarchical cortical route improves model performance on a range of tasks by enabling rapid responses to simple stimuli while preserving deep processing for complex decisions.
Key Findings
- The Shallow Brain Hypothesis: Extending a 2023 proposal, the team demonstrates that the brain does not rely exclusively on step-by-step hierarchical processing. Instead, it uses parallel interactions between deep cortical layers and shallower subcortical circuits.
- Functional Complementarity: Experiments on decision-making tasks reveal complementary roles: subcortical routes support fast stimulus-response behaviors, whereas cortical hierarchies handle intricate problem-solving and context-dependent decisions.
- Efficiency Gains: The parallel architecture permits more flexible allocation of computational resources. For simple tasks, a shallow route suffices—deep learning architectures can be unnecessarily heavy for those cases.
- Greater Biological Realism: Unlike most artificial neural networks, the proposed model incorporates the feedback loops and short, direct pathways present in real neuroanatomy, bringing computational models closer to the organization of the human nervous system.
Source: EBRAINS
A research team in the Netherlands has introduced a computational design that merges cortical hierarchical processing with parallel subcortical pathways—an approach that could reshape how neuroscientific insights influence future AI development.
Conventional deep learning systems typically route information through many sequential layers that model cortico-cortical connectivity. While effective for complex recognition and inference, these architectures underrepresent non-hierarchical brain components: subcortical networks and the bidirectional interactions between cortical and subcortical areas. Those deeper structures are central to movement, emotion, and fast stimulus-response learning.
Supported by the Human Brain Project and published in Current Research in Neurobiology, the study describes a computational model that explicitly incorporates cortico-subcortical connectivity. The resulting architecture features two parallel processing streams: a deep, hierarchical cortical route and a fast, shallow subcortical route. This dual-pathway design better reflects known anatomy and functional divisions of the brain.
To demonstrate practical value, the authors implemented the cortical component in two widely used frameworks: a convolutional feedforward neural network and a hierarchical predictive coding model. Both variants, when augmented with the shallow subcortical pathway, reproduced behavioral patterns observed in humans and nonhuman primates performing a perceptual, context-dependent decision-making task.
Their results indicate that the shallow subcortical route leads decisions on easy trials where rapid, stimulus-driven responses are advantageous, whereas the deeper cortical network becomes necessary as task complexity increases. This division of labor yields faster responses when appropriate and preserves the capacity for detailed inference when required.
Overall, the parallel cortico-subcortical architecture offers a principled, biologically grounded alternative to strictly hierarchical models, with implications for efficient AI design and a deeper understanding of brain computation.
Key Questions Answered
A: Deep networks excel at complex pattern recognition but can be excessive for immediate, reflexive responses. A shallow subcortical route provides a biological shortcut that enables near-instant reactions to simple stimuli without routing information through many deep layers.
A: Many AI models are inspired mainly by cortical hierarchies and therefore overlook subcortical structures involved in emotion, survival behavior, and basic learning. Incorporating these pathways can make models faster, more efficient, and more adaptable.
A: By enabling both rapid, instinctive responses and deliberate, analytical processing, the two-route architecture mirrors how humans balance gut reactions with reflective thought, improving behavioral flexibility and prioritization.
Editorial Notes
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full by the editorial team.
- Additional context was added by staff to clarify implications for AI and neuroscience.
About this neuroscience and AI research news
Author: Helen Mendes Lima
Source: EBRAINS
Contact: Helen Mendes Lima – EBRAINS
Image: Credit to Neuroscience News
Original Research: Open access. “A computational architecture incorporating shallow brain networks: integrating parallel cortical and subcortical processing” by Kwangjun Lee, Lorenzo Gabriele Baracco, Cyriel M.A. Pennartz, Mototaka Suzuki, and Jorge F. Mejias. DOI: 10.1016/j.crneur.2026.100155
Abstract
A computational architecture incorporating shallow brain networks: integrating parallel cortical and subcortical processing
Most artificial neural networks emphasize deep hierarchical structures modeled after cortico-cortical connectivity. Non-hierarchical brain features—particularly subcortical pathways and cortico-subcortical interactions—are underrepresented despite their computational importance. Motivated by neuroanatomical evidence, the authors present a computational model that combines cortical hierarchies with parallel subcortical processing.
They implement the cortical hierarchy in two alternative ways: a convolutional feedforward network and a predictive coding framework. Both variants, when coupled with a shallow subcortical pathway, replicate behavioral data from perceptual, context-dependent decision tasks in humans and monkeys. The model shows subcortical circuits leading on easier trials and the cortical hierarchy governing harder ones, suggesting parallel cortico-subcortical processing is a fundamental computational principle that should inform both neuroscience and AI design.