Hypothesis
Deep, nonlinear models would predict age and sex from structural and functional brain scans better than linear models as training samples grow, as they do on standard machine-learning image benchmarks.
What failed
On MNIST and Fashion benchmarks, accuracy rose from linear to shallow to deep models and with sample size; on UK Biobank brain scans, simple linear models performed on par with kernel and deep models at every sample size up to about 10,000, and kept improving.
Broken assumption
Deep learning's advantage on natural images carries over to predicting phenotypes from brain scans at today's sample sizes.
Why
The authors conclude that the nonlinearities in common brain scans remained inaccessible to both kernel and deep methods at any scale examined, while linear-model gains had not yet saturated.
From the published paper: Schulz et al. Nature Communications, 2020. ↗CC BY