DEBI-NN Study: Genetic Algorithm Beats Gradient Descent for Small Medical Datasets

A new arXiv preprint (2605.27411) by Boukhari et al. tests whether gradient descent (GD) can effectively train DEBI-NN, a recently proposed architecture that encodes connection weights via Euclidean distances between neurons in a spatial map. Unlike classic neural networks that directly optimize each weight, DEBI-NN dramatically reduces trainable parameters but traditionally relies on a genetic algorithm (GA) for learning. The team designed a custom spatial backpropagation scheme to enable GD on DEBI-NN and ran classification benchmarks on four datasets: a synthetic non-linear "two-moons" problem, two clinical radiomics datasets (DLBCL, HECKTOR), and a fetal cardiotocography set, with sample sizes ranging from 85 to 2,126.

The results are decisive: GA outperformed GD on every metric. On synthetic data, GA hit 100% accuracy while GD stalled at 83%. On DLBCL: 83% vs 78%; HECKTOR: 80% vs 67%; fetal: 81% vs 66%. The authors attribute GD's failure to "entangled gradients" caused by high neuron interdependence in DEBI-NN's spatial encoding, which undermines classic backpropagation. They conclude that evolutionary strategies remain the appropriate optimizer for this architecture, especially in low-data medical settings where robust decision boundaries are critical.