USC lecture notes derive closed-form AUC scaling laws and a 'spectral horizon' for finite-data ML
PTenigma · x · 2026-09-16
Paul Thompson (USC) published open lecture notes extending his zeta law of discoverability into a spectral theory of learning curves for finite-data ML: recoverable eigenmodes define a spectral horizon K(N), power-law spectra yield truncated zeta-sum closed-form scaling laws for AUC(N), and a martingale/Malliavin extension gives confidence cones for future performance — predicting model crossovers, staircase curves, and when more data or new sensing modalities pay off most.
More from Research
- Is mathematical abstraction just human assembly code? The case that AI can grind through what we can't — Jsevillamol · 2026-09-16
- Rumors: OpenAI may have cracked a second Millennium Problem, holding major announcements amid Navier-Stokes backlash — TheGoldenLeaper · 2026-09-16
- University of Manchester forecasts UK air pollution with NVIDIA Earth-2 generative models — nordicinst · 2026-09-16
- AI designs drugs in months, but India takes a year to approve trials while Australia and China take weeks — jajoosam · 2026-09-16
- CTD Commons aims to publish regulatory files from failed drug programs so future teams can learn from them — alex_peys · 2026-09-16
- Single-cell mass spec (plexDIA) quantifies more gene products per cell than RNA measurements — iskander · 2026-09-16