Research
Postdoc at KAIST AMILab. Designing brain-inspired cognitive architectures for systematic generalization and continual learning.
My work sits at the intersection of cognitive neuroscience and AI. The central question driving everything: what computational principles does biological intelligence use to generalize so gracefully, and can we build AI systems around those same principles?
I approach this through a two-direction loop:
- AI for Neuroscience β using computer vision and generative models as βpractical microscopesβ to decode neural and behavioral data that was previously too complex to quantify.
- Neuroscience for AI β drawing on canonical neural computations (grid cells, cortical columns, complementary learning systems) to architect AI that generalizes more robustly.
Core Principles
Three mechanistic ideas from neuroscience anchor everything I build:
- Universal Reference Frames β Abstract knowledge anchored to stable spatial representations (inspired by hippocampal grid cells and the Thousand Brains theory).
- Predictive Modeling in Canonical Circuits β World models learned through local prediction, mirroring cortical column computations.
- Structure / Content Factorization β Separating reusable structure (the βgrammarβ) from variable content (the βwordsβ) to enable compositional generalization and lifelong learning without catastrophic forgetting.
4-Stage Cognitive Architecture
These principles map onto a modular hierarchy:
| Stage | Name | Function |
|---|---|---|
| I | Object-Centric Perception | Grounded object representations via Slot Attention + grid-cell reference frames |
| II | Predictive Abstraction | JEPA (Joint Embedding Predictive Architecture)-style prediction β discrete symbol conversion |
| III | Semantic Consolidation | Episodic β semantic knowledge integration (CLS theory) |
| IV | Metacognitive Control | MoE-PRM (Mixture-of-Experts Process Reward Model): dynamic routing between System 1 intuition and System 2 reasoning |
Research Themes
These three pillars integrate the core principles above with my research trajectory.
Structured Representation & Memory Consolidation
Compositional generalization through structure / content factorization; context-sensitive coordination of working and long-term memory; episodic-to-semantic integration via complementary learning systems (CLS); representational hierarchies inspired by cortical columns; sparse and disentangled coding for continual learning.
Multi-modal Grounding via Reference Frames
Spatial anchoring through grid-cell-inspired coding; cross-modal binding of vision, language, and other sensory streams; world models learned through local prediction in canonical circuits; reference-frameβbased generalization.
Social & Context-Adaptive Cognition
Empathy and social inference grounded in perceptionβaction coupling and theory of mind; context-conditioned representations that modulate behavior across situations; multi-agent interaction and active inference. This pillar grew out of rodent affective-empathy neuroscience and now informs human-aligned multi-modal AI.
Publications and CV: Google Scholar Β· CV
