Granularity and the 'birdiest bird': self-supervised clusters cut across ImageNet labels

y_m_asano · x · 2026-09-09

Researcher Yuki Asano reflects on Serge Belongie's workshop talk on granularity, tying it back to his Oxford VGG self-supervised clustering work:

• Two ways to define granularity — Venn-diagram boundaries vs distance to the 'birdiest' bird — expose the fuzziness of category edges.
• The Self-Labelling method learns features and pseudo-labels simultaneously by optimizing cross-entropy while maximizing information, generating labels for any image dataset without annotation.
• Notable empirical finding: highly coherent self-supervised clusters often span multiple ImageNet classes (e.g. 9 beach images from 3 different classes), showing unsupervised structure doesn't align with human label taxonomies.

Original post →

More from Research

• NumeriaPlus: 300-hour egocentric activity dataset with dense ground truth for world models — ducha_aiki · 2026-09-09
• OpenAI claims 10,000 agents proved Navier-Stokes blowup in 88 hours; Clay Institute won't accept it — AdaptiveAgents · 2026-09-09
• LLM Zelda Level Generator Hits Diversity Metric but Produces Visually Boring Rooms — Amidos2006 · 2026-09-09
• Honest limitation of an evolutionary Zelda generator: high fitness ≠ visual diversity — Amidos2006 · 2026-09-09
• LLM-evolved global functions are readable: a ~5-line counting function is most common — Amidos2006 · 2026-09-09
• One global function takes LLM-evolved level generation from ~0% to ~100% playability — Amidos2006 · 2026-09-09

Research channel →