Sparse Representation
We study how information can be represented through sparse active patterns rather than only dense embeddings. Sparse representations may support interpretability, compositionality, and more efficient state formation.
Predictive Coding
We study predictive coding as a learning direction where systems update through prediction error and local learning signals, rather than relying only on global backpropagation.
Relational State
We study how AI systems can bind entities, roles, events, and context into reusable internal state.
Memory & Continual Learning
We study how systems can retain, update, forget, and reuse information across time without simply replaying all prior context.
Neuroscience-Inspired Learning
We use neuroscience as inspiration for architectural constraints, not as a claim that our systems replicate the brain.
Multimodal Grounding
Future work explores how sparse and predictive systems can extend across language, vision, audio, code, and agents.
Human-AI Interaction
We study how better memory, state, and context can support more natural interaction between humans and AI systems.