GRINCO: Group-invariant coresets cut redundant labeling by selecting orbits
New method avoids wasting budget on rotated or transformed duplicates of the same image.
Active learning helps reduce labeling costs by querying the most informative samples, but standard coreset methods treat each data point independently and often waste budget on transformed versions of the same instance—like two rotated copies of the same image. To solve this, the authors introduce GRINCO (Group-INvariant COreset), a framework that performs sample selection in the quotient space induced by a transformation group. This means acquisition operates on entire orbits (sets of all possible transformations of a sample) rather than raw samples.
GRINCO offers two practical approaches to define a quotient metric: using canonical representatives (e.g., normalizing rotation before selection) or learned orbit-separating invariant embeddings. It then combines quotient-space k-center clustering with an orbit-averaged training loss. The paper also derives a generalization bound that relates excess orbit-averaged risk to quotient-space coverage, label uncertainty, and intra-orbit variability. Experiments on synthetic scale-invariant data and image benchmarks with rotation-induced redundancy show that GRINCO significantly improves orbit coverage and label efficiency over conventional coreset baselines, especially when group redundancies are substantial.
- GRINCO selects samples from orbits (sets of transformed versions) rather than raw samples, avoiding redundant labeling
- Uses either canonical representatives or learned orbit-separating invariant embeddings for practical quotient metrics
- Experiments show improved label efficiency on rotation-heavy image benchmarks compared to standard coreset methods
Why It Matters
For practitioners, GRINCO slashes labeling costs by avoiding redundant queries, especially in domains with known symmetries like vision.