Explaining Modular Representations with Range
An ICLR 2025 paper proposing that range, not just statistical independence, is the key factor that determines whether neural representations are modular or mixed.
About NABI
Natural and Artificial Brain Intelligence
Based in Seoul National University, College of Medicine
NABI is a research group which aims to translate clinical experience into insights through computational neuroscience. We also aim to leverage real-life data to advance our understanding of the human brain. Ultimately, we aim to develop computational models that can be leveraged in clinical or other real-life applications.
Focused on the latest topics in deep learning, computational neuroscience, and reinforcement learning
We analyze papers and implement code on the latest topics from prominent journals and conferences in deep learning, computational neuroscience, and reinforcement learning.
A multidisciplinary community
Students from various fields, including medicine, biology, and computer science, come together to develop interdisciplinary skills. We hold an end-of-year symposium and a presentation at Seoul National University Hospital so various members can present their research and hold discussions with peers from various fields.
Research Highlights
Key Papers reviewed at NeuroAI
Groundbreaking studies at the intersection of neuroscience and artificial intelligence
Could a neuroscientist understand a microprocessor?
Neuroscience techniques, when applied to a microprocessor demonstrate even human-designed systems may be difficult to understand.
Meta-reinforcement learning via orbitofrontal cortex
The hypothesis that the OFC (orbitofrontal cortex) plays a role in meta-reinforcement learning is proposed, and two distinct algorithms with distinct timescales are proposed.
Transformer as a hippocampal memory model
Based on NMDAR nonlinearity, a circuit model of memory consolidation is proposed, establishing a link between the hippocampus and the transformer architecture.
Prefrontal cortex as a meta-reinforcement learning system
The prefrontal cortex operates as a flexible free-standing system through a process of meta-reinforcement learning.
Human cognition and neuromodulatory integration
By identifying the signatures of system-wide neuromodulatory integration, we advance our knowledge the functional organization of the human brain.
Scaffolding cooperation with deep RL
A novel AI-driven social planner dynamically reshapes human interaction networks to promote cooperation by welcoming defectors into collaborative groups rather than isolating them.
A recurrent planning model explains hippocampal replay
A recurrent neural network model demonstrates how the prefrontal cortex orchestrates planning by sampling imagined action sequences, termed rollouts.
Brain-wide behavior representations in C. elegans
Whole-brain calcium imaging and probabilistic modeling is used to map how distinct neuron classes in C. elegans encode behavior across various states and timescales.
Emergence of syntax in a cerebellar circuit
A biologically-constrained artificial neural network modeled after cerebellar circuity demonstrates a single recurrent circuit can perform word prediction and syntactic structure recognition
Read more about our research
View all posts »Research notes from our monthly meetings
Time Encoding in the Lateral Entorhinal Cortex
A presentation of the 2025 Science paper shows how the Lateral Entorhinal Cortex (LEC) encodes time via two mechanisms. A continuous 'slow drift' and abrupt 'shifts' at event boundaries.
Compositional Hippocampus Constructs Future Behavior
Hippocampal memory is compositional, and performs consolidation primarily through replay, allowing for zero-shot generalization and the construction of future behavior
Vector Hash Memory Model
The VectorHASH model separates memory content from a fixed, stable scaffold derived from hippocampal structure, integrating episodic and spatial memory