Summary: Current AI models mimic the brain’s cortex, the “high-level” outer layer, but they almost entirely ignore the ancient, deep structures beneath it. A research team has proposed a new computational architecture that integrates these subcortical structures.
The study demonstrates that adding a “fast, shallow” subcortical route alongside the “deep, hierarchical” cortical route makes computer models more flexible, efficient, and biologically plausible.
Key Findings
- The Shallow Brain Hypothesis: Building on their 2023 theory, the team proved that the brain does not rely solely on step-by-step hierarchical layers. Instead, it uses parallel interactions between deep and surface regions.
- Functional Complementarity: Tests on decision-making tasks showed the two pathways work in harmony. The subcortical route guides quick reactions, while the cortical network takes over for intricate problem-solving.
- Efficiency Gains: This parallel architecture allows AI to process information more flexibly, suggesting that “deep learning” may actually be too deep for simple tasks, wasting computational resources.
- Biological Realism: Most artificial neural networks lack the feedback loops and “shortcuts” found in the human brain. This model brings AI hardware and software closer to the actual anatomy of the human nervous system.
Source: Ebrains
A team of researchers in the Netherlands has proposed a new way of designing computer models of the brain, an approach that could also influence future artificial intelligence (AI) systems.
In most deep learning architectures, information is processed step by step through tens of layers inside the cortex, the brain’s main structure involved in high-level functions like perception and decision-making.
However, neuroscientists know that the cortex is also closely connected with deeper brain regions, known as subcortical structures – which are involved in processes such as regulating body movement, emotion and learning stimulus-response behaviours – and these connections are overlooked by most artificial neural networks.
In a new study supported by the Human Brain Project and published in Current Research in Neurobiology, the researchers introduce a computational model that incorporates these connections, combining a hierarchical architecture typical of the cortex with faster, subcortical pathways. This proposed architecture is more parallel – having a hierarchical, cortical route and a “shallow”, subcortical route – and may better reflect how the brain works.
“Our model addresses key limitations in existing deep learning and predictive coding networks, offering a more biologically plausible and functionally advantageous alternative”, say the authors.
The work builds on the authors’ 2023 proposal of the “Shallow Brain Hypothesis”, which argues that the brain relies on both hierarchical processing in the cortex and parallel interactions with subcortical regions. The team has now developed a model combining both pathways found in the brain.
They implemented this approach using two common AI frameworks – a convolutional neural network and a hierarchical predictive coding model – and tested it on a decision-making task. Their results show that the two pathways complement each other: the fast subcortical route can guide simple stimulus-response decisions, while more complex tasks rely on the ‘deep’ cortical network.
Together, this parallel architecture allows for more flexible and efficient processing, suggesting that current AI models may well be missing an important principle of how the brain works.
Key Questions Answered:
A: Deep learning is great for recognizing a face in a crowd, but it’s overkill for pulling your hand away from a hot stove. The “Shallow” route provides a biological shortcut. It allows an AI (or a brain) to react instantly to simple stimuli without waiting for the data to travel through dozens of complex layers.
A: Yes. Most AI is modeled exclusively on the cortex. This study argues that by ignoring subcortical structures, the parts of the brain that handle emotion, survival instincts, and basic learning, we are building AI that is smart but inflexible and inefficient.
A: It introduces a sense of priority. A “Shallow Brain” AI can have “gut reactions” for simple tasks while “thinking deeply” for others. This mirrors how humans actually function, balancing instinct with intellect.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context added by our staff.
About this neuroscience and AI research news
Author: Helen Mendes Lima
Source: EBRAINS
Contact: Helen Mendes Lima – EBRAINS
Image: The image is credited 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. Current Research in Neurobiology
DOI:10.1016/j.crneur.2026.100155
Abstract
A computational architecture incorporating shallow brain networks: integrating parallel cortical and subcortical processing
Artificial neural networks commonly have deep hierarchical structures that were originally inspired by the neuroanatomical evidence of cortico-cortical connectivity pattern found in the mammalian brain.
Largely under-represented in those models are non-hierarchical aspects of brain architecture, namely the subcortical pathways and the interactions between cortical and subcortical areas regardless of their hierarchical locations.
Inspired by this principle, we present a computational model combining cortical hierarchical processing with subcortical pathways based on neuroanatomical evidence.
We show the versatility of our model by implementing the cortical hierarchy in two alternative ways—a convolutional feedforward network and a predictive coding network.
Both model variants can replicate behavioral observations in humans and monkeys on a perceptual context-dependent decision-making task.
The model also reveals that subcortical structures lead decisions for easy trials while the more complex hierarchical network is necessary for the harder trials.
Our results suggest that the parallel cortico-subcortical processing explored in the model represents a fundamental property that cannot be neglected in understanding the computational principles used by the brain.
