JEM trained in ~20 H100 hours nearly matches fine-tuned DINOv3 on plant features
randall_balestr · x · 2026-11-10
- The author finished training JEM for 100 epochs on a 250k-image plant dataset using just over 20 H100 hours, landing only 1 point behind a heavily fine-tuned supervised DINOv3 (Pl@ntNet checkpoint).
- Patch features were visualized with PCA→RGB, comparing LeJEPA, JEM with and without LayerScale, and supervised DINOv3 across common and rare plant classes.
- Notably, LeJEPA performs worse yet produces crisp features — it only gets a loss on its global embedding, never on patch tokens.
- Removing LayerScale from JEM was the right call: smoother maps that cleanly segment leaves and flowers.
- Rare classes remain the main weakness (likely a sampling issue), and the LeJEPA baseline is still being tuned.
More from Research
- Researchers debate optimizer equivariance: argument without orthogonality is more convincing — jasondeanlee · 2026-11-10
- Claude Fable 5.2 tops VIALS-AA science visual reasoning benchmark at 67% with heavy tool use — ArtificialAnlys · 2026-11-10
- WVU researcher presents 10 years of tractometry work on Alzheimer's white matter vulnerability — PTenigma · 2026-11-10
- New cryo-EM structure: pSTAT1 bound to importin alpha 5 (PDB 11GJ) — chaitjo · 2026-11-10
- ARLoop: fully-open 1.2B looped transformer beats Huginn-3.5B and Qwen3.5-2B on all 8 benchmarks — hhsun1 · 2026-11-10
- RAGCap-Bench: A Fine-Grained Benchmark for Intermediate Capabilities in Agentic RAG — kalyan_kpl · 2026-11-10