Towards Algorithmic Analytics for Large-scale Datasets

Danilo Bzdok (1, 2, 3) , Thomas E. Nichols (4) , Stephen Smith (5)

1 Department of Psychiatry, Psychotherapy and Psychosomatics [Aachen]
2 NEUROSPIN - Service NEUROSPIN
3 PARIETAL - Modelling brain structure, function and variability based on high-field MRI data
4 WIN - FMRIB - Wellcome Trust Centre for Integrative Neuroimaging
5 Big Data Institute

The traditional goals of quantitative analytics cherish simple, transparent models to generate explainable insights. Large-scale data acquisition, enabled for instance by brain scanning and genomic profiling with microarray-type techniques, has prompted a wave of statistical inventions and innovative applications. Modern analysis approaches 1) tame large variable arrays capitalizing on regularization and dimensionality-reduction strategies, 2) are increasingly backed up by empirical model validations rather than justified by mathematical proofs, 3) will compare against and build on open data and consortium repositories, as well as 4) often embrace more elaborate, less interpretable models in order to maximize prediction accuracy. Here we review these trends in learning from "big data" and illustrate examples from imaging neuroscience.

Mots clés: machine learning; data science; open science; reproducibility; explainable AI; deep phenotyping

Domaines: Bio-informatique [q-bio.QM]; Intelligence artificielle [cs.AI]

Danilo Bzdok, Thomas E. Nichols, Stephen Smith. Towards Algorithmic Analytics for Large-scale Datasets. Nature Machine Intelligence, 2019, ⟨10.1038/s42256-019-0069-5⟩. ⟨hal-02178410⟩