An evolutionary tree for machine learning's unknown classes

What it does This repository is a curated bibliography of papers tackling Novel Class Discovery (NCD)—the machine learning problem of spotting new categories in unlabeled data when you already have labeled examples from different, known classes. Unlike a generic arXiv dump, it organizes surveys, preprints, and peer-reviewed work into a taxonomy of problem settings, from core NCD to Generalized Category Discovery and Federated GCD. The maintainer also provides an evolutionary tree diagram that visualizes how these research threads have diverged over time.

The interesting bit The list doubles as a field map: it includes an “evolutionary tree” diagram and explicitly flags not-yet-codified directions like Incremental GCD and Semantic Category Discovery as “TODO,” treating the bibliography as a living draft of the field rather than a static index.

Key highlights

• Covers core NCD and its offshoots: open-world semi-supervised learning, continual discovery, class-incremental settings, and multi-modal variants.
• Separates survey papers, preprints, and conference publications (CVPR, ICLR, ICML) with direct links to papers and code where available.
• Provides an evolutionary tree diagram (assets/ncd-tree.png) showing how problem settings branched.
• Explicitly acknowledges itself as a non-exhaustive list, leaving headroom for emerging variants.
• Many entries include direct links to both arXiv PDFs and official implementation repositories.

Caveats

• The README offers minimal commentary or synthesis; it is a reference list, not a tutorial or benchmark comparison.
• Several proposed problem settings (e.g., IGCD, SCD) are listed as “TODO,” so the taxonomy is still incomplete.
• The list is explicitly non-exhaustive, so you may still need to search adjacent keywords.

Verdict Grab this if you are a researcher or practitioner trying to navigate the sprawling NCD literature and locate the right variant for your semi-supervised use case. Skip it if you are hunting for a unified code framework or ready-to-run benchmarks.