AI for Science Warning: Explained Variance Is Not Explained Science in PCA
bravo_abad · x · 2026-09-23
The author argues that the first principal component of PCA is merely the direction of greatest variance, not necessarily the most scientifically meaningful one, using spectral data as an example: an intense peak that barely varies across samples contributes little to the leading components, while a weaker spectral region that varies strongly can dominate them.
- The dominant source of variation may not be the phenomenon of interest at all, but baseline drift, sample thickness, illumination, temperature fluctuations, batch effects, or instrumental artifacts; PCA cannot tell which variation is scientifically meaningful.
- A single physical effect need not land on a single component: when two components explain similar variance, PCA cannot separate them, so one effect can spread across several components or several effects blend into one.
- Interpreting PCA therefore requires going back from components to original variables and asking what physical, chemical, biological, or experimental variation each component represents; preprocessing (standardizing, normalizing spectral regions, rescaling) changes the geometry and thus the components.
- Conclusion: scientific datasets are often highly redundant, so a few components can retain most of the variance, but "explained variance is not explained science." The idea comes from the author's book Inteligencia artificial para estudiantes de ciencias.
Related event: New AI for Science Book Warns PCA Components Aren't Always Meaningful(2 posts)→
More from AGI Musings
- Politico poll: more people say AI helped than hurt them in every country surveyed — neil_chilson · 2026-09-23
- Rewatching 2001: today's AI lies without blushing — and apologizes like a pro — sankin_eth · 2026-09-23
- Researcher on retraining the 'idea muscle' as AI commoditizes intelligence — dejavucoder · 2026-09-23
- Ethan Mollick: has a frontier model actually caused a major security incident in normal deployment? — emollick · 2026-09-23
- Scholar: Refusing AI in class and 'meat computation' hype are both bad pedagogy — ruthstarkman · 2026-09-23
- Dev Slams 'Swarm' Narrative: Trillion-Dollar Valuations Built on a Sampler With No Brakes — gerardsans · 2026-09-23