summary: Current AI models mimic the brain’s cortex, or “high-level” outer layer, but almost completely ignore the ancient deep structures beneath. The research team proposed a new computational architecture that integrates these subcortical structures.
This study shows that adding a “fast and shallow” subcortical route in addition to a “deep and hierarchical” cortical route makes computational models more flexible, efficient, and biologically plausible.
Main findings
- Shallow brain hypothesis: Based on the 2023 theory, the research team has proven that the brain does not rely solely on progressive hierarchical layers. Instead, we use parallel interactions between deep and surface regions.
- Complementarity of functions: Tests on decision-making tasks showed that the two pathways work in harmony. Subcortical routes lead to quick reactions, and cortical networks take over complex problem solving.
- Increased efficiency: This parallel architecture allows AI to process information more flexibly, suggesting that “deep learning” may actually be: too much It is deep for simple tasks and wastes computational resources.
- Biological realism: Most artificial neural networks lack the feedback loops or “shortcuts” found in the human brain. This model brings AI hardware and software closer to the actual structure of the human nervous system.
sauce: Ebrains
A team of Dutch researchers has proposed a new way to design computer models of the brain. This approach could also impact future artificial intelligence (AI) systems.
In most deep learning architectures, information is processed step by step through dozens of layers within the cortex. The cortex is a major structure of the brain involved in high-level functions such as perception and decision-making.
However, neuroscientists know that the cortex is also closely connected to deeper brain regions known as subcortical structures that are involved in processes such as body movement, emotional regulation, and stimulus-response behavior learning, but these connections are overlooked in most artificial neural networks.
A new study supported by the Human Brain Project and published in the journal Current research in neurobiologyThe researchers introduced a computational model that incorporates these connections, combining a hierarchical structure typical of the cortex with faster subcortical pathways. This proposed architecture is more parallel, has a hierarchical cortical route and a “shallow” subcortical route, and may better reflect how the brain works.
“Our model addresses key limitations of existing deep learning and predictive coding networks and provides a more biologically plausible and functionally advantageous alternative,” the authors state.
The study builds on the authors’ 2023 proposal for the “shallow brain hypothesis,” which argues that the brain relies on both hierarchical processing in the cortex and parallel interactions with subcortical regions. The research team has now developed a model that combines both pathways found in the brain.
They implemented this approach using two popular AI frameworks: convolutional neural networks and hierarchical predictive coding models, and tested it on a decision-making task. Their results show that the two pathways complement each other. Fast subcortical pathways can guide simple stimulus-response decisions, but more complex tasks rely on “deep” cortical networks.
This parallel architecture allows for more flexible and efficient processing, suggesting that current AI models may be missing important principles about how the brain works.
Answers to key questions:
answer: Deep learning is great for recognizing faces in a crowd, but it’s too much for keeping your hands off a hot stove. The “shallow” route provides a biological shortcut. This allows the AI (or brain) to respond instantly to simple stimuli without having to wait for the data to pass through dozens of complex layers.
answer: yes. Most AI models only the cortex. The study argues that by ignoring subcortical structures, the part of the brain that deals with emotions, survival instincts, and basic learning, we are building intelligent but inflexible and inefficient AI.
answer: It brings a sense of priority. A “shallow brain” AI can have “intuitive reactions” to simple tasks and “deep thinking” to other tasks. This reflects how humans actually function, balancing instinct and intelligence.
Editorial note:
- This article was edited by the editors of Neuroscience News.
- Journal articles were reviewed in full text.
- Additional context added by staff.
About this neuroscience and AI research news
author: helen mendez lima
sauce: Ebrains
contact: Helen Mendez Lima – EBRAINS
image: Image credited to Neuroscience News
Original research: Open access.
“Computational Architecture Incorporating Shallow Brain Networks: Integrating Parallel Cortical and Subcortical Processing” by Kwangjun Lee, Lorenzo Gabriele Baracco, Cyriel MA Pennartz, Mototaka Suzuki, and Jorge F. Mejias. Current research in neurobiology
DOI:10.1016/j.crneur.2026.100155
abstract
Computational architecture incorporating shallow brain networks: integrating cortical and subcortical parallelism
Artificial neural networks generally have deep hierarchical structures, which were originally inspired by neuroanatomical evidence of cortico-cortical connectivity patterns found in the mammalian brain.
These models largely underestimate non-hierarchical aspects of brain structure, that is, subcortical pathways and interactions between cortical and subcortical regions regardless of hierarchical position.
Inspired by this principle, we present a computational model that combines cortical hierarchical processing and subcortical pathways based on neuroanatomical evidence.
We demonstrate 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 reproduce behavioral observations in humans and monkeys regarding perceptual context-dependent decision-making tasks.
The model also revealed that subcortical structures drive decision-making during easy trials, whereas more complex hierarchical networks are required for more difficult trials.
Our results suggest that the cortical and subcortical parallelism investigated in our model represents a fundamental property that cannot be ignored in understanding the computational principles used in the brain.
