Reproduce “Tree of Concepts: Interpretable Continual Learners in Non-Stationary Clinical Domains” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongkyu Cho, Xiyue Li, Samrachana Adhikari, Rumi Chunara
A framework called "Tree of Concepts" enables interpretable continual learning by combining a frozen decision tree to define stable, rule-based concepts with an adaptive concept bottleneck model. This approach consistently demonstrates an improved balance between learning new information and retaining old knowledge, while maintaining auditable concept consistency across various non-stationary tabular healthcare datasets.
Key insights
- Decoupling the stable definition of an explanation interface from the adaptive learning of predictive representations effectively reconciles the tension between interpretability and continual learning.
- Utilizing a fixed, rule-based concept vocabulary from a frozen decision tree ensures that explanation semantics remain consistent and auditable across sequential model updates.
- This architecture, coupled with replay, yields a superior trade-off between model plasticity (adaptation) and stability (knowledge retention) in dynamic tabular data environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.