The abstract is read at the publisher; the summary is JClub's.
Gene co-expression–initialized sparse multilayer perceptrons classify handwritten digits with 65% accuracy using 100 training examples, far above a dense baseline at 11%.
- Why it matters: It is unclear whether sparse, modular, small-world biological network structure can serve as an inductive bias for data-efficient sparse learning.
- What they did: The study built sparse multilayer perceptrons whose adjacency was inferred via mutual information from molecular, structural, functional, and behavioral biological networks, then benchmarked four classification tasks against dense and density-matched synthetic graphs.
- The result: Biologically initialized connection patterns—not edge counts—produced a low-data advantage and lower trial-to-trial variance that rewired or degree/scale-free/small-world matched graphs did not replicate.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.