Study finds data scaling in scientific AI is not always smooth
bravo_abad · x · 2026-08-19
A study on neural networks learning solid optical properties reveals that while model size follows standard power laws, scaling with dataset size behaves differently. Researchers observed non-smooth improvements or plateaus across architectures, challenging the assumption that more data always yields smooth gains in scientific AI.
More from Research
- Tech opinion: Pre-training transfer learning weaker than expected 5 years ago — jack_w_rae · 2026-08-19
- MIT develops mathematical framework for bio-inspired adaptive materials — GioeleZardini · 2026-08-19
- Ling-3.0 Open Source: 6 Base Checkpoints Released with WSM Training Method — bclavie · 2026-08-19
- Study: LMs predict forward reading but struggle to explain regressions — tallinzen · 2026-08-19
- Nature Medicine cover: AIDO vision calls for multiscale "digital organisms" beyond AlphaFold — HongyiWang10 · 2026-08-19
- How 'models' were quantised in the 19th century: British Library dataset — vanstriendaniel · 2026-08-19