New Study Reveals Hidden Bias in Brain and AI Measurements
Your brain and AI measurements might be misleading—here's why it matters for health and tech.
A neural population’s measured dimensionality is not simply the code: the slope of the participation ratio against subsample size, often reported for cortical populations and language-model layers as a number belonging to the system, does not belong to the system. One patch of 11,000 mouse V1 neurons returns three slopes—0.25 along random stimulus subsets, 0.31 along drift direction, and 0.35 along neuron count—and only the middle one carries structure; the other two are the sampling floor.
The study splits every exponent into that floor plus a shift earned along a declared axis, then reads the shift against two nulls: permuting labels, and permuting labels only among samples that share a nuisance. On V1, the direction-aligned shift clears the label null in 8 of 8 grating recordings, is positive
- A common measurement in brain and AI research is biased by the measurement process itself.
- The new method separates true signals from noise, revealing hidden patterns in both mouse brains and language models.
- This could lead to more reliable neuroscience studies and better AI systems.
Why It Matters
Better measurements mean more trustworthy science, which could speed up treatments for brain disorders and improve AI reliability.