Should rare classes be merged into an "Other" bucket or treated as OOD detection?
neonhexe · reddit · 2026-08-20
A Reddit ML discussion: when training a dog-breed image classifier, common breeds have plenty of samples while long-tail breeds have only a handful. Is grouping them into a catch-all "Other breed" class harmful? The asker's intuition is that it forces weirdly-shaped hyperplanes separating very different dogs (chihuahuas vs wolf-like dogs) in latent space. They suggest instead treating underrepresented breeds as an out-of-distribution detection problem — keep only well-represented classes in training and drop the rest — and ask for research pointers.
More from Research
- Paper suggests tuning hyperparameters on small scales before large scale adaptation — JJitsev · 2026-08-20
- GenBio AI releases AIDO Cell, a world model of the human cell — TinfoilTricorn · 2026-08-20
- Robotics in-context learning sparks debate over definition of gradient updates — NielsRogge · 2026-08-20
- SkillGate: Fixing Credit Starvation in Long-Horizon Agent Skill Selection — SJTU · 2026-08-20
- AdaPop: Adaptive Popularity for LLM Unlearning Reduces Leakage — Anna Borisiuk · 2026-08-20
- Clarifying Hyperparameter Claims in Scaling Laws Research — JJitsev · 2026-08-20