Deep Genetic Folding: An RNA-Inspired Hybrid Architecture For Image Classification

Abstract

Genetic Folding (GF) was introduced in 2011 as an evolutionary algorithm inspired by the way an RNA molecule folds onto itself. The original metaphor pairs complementary bases. Most later GF variants kept the parse tree but dropped the literal base-pairing rule. In this study we restore that biology. We propose Deep Genetic Folding (DGF), a hybrid architecture for image classification. The chromosome is a nucleotide string over {A, U, G, C}. Folding is performed by the Nussinov maximum-pairing dynamic program. The folded secondary structure becomes the architecture of a layered feature extractor. Each base pair contributes one feature column. Nested pairs form deeper layers. A small Random Fourier feature bank and a two-block residual MLP head close the gap to strong gradient baselines. The top-three chromosomes from the final population vote at inference. We evaluate DGF on Fashion-MNIST downsampled to 14 x 14 across five seeds under a strict no-leakage 2500 / 500 / 1000 protocol. DGF reaches 85.22% mean test accuracy with a 1.07% standard deviation. It ranks first on four of five seeds and achieves the best mean rank of 1.20. It beats Support Vector Machines with an RBF kernel by 0.50 percentage points, Multi-Layer Perceptrons by 1.86 pp, Random Forest by 2.06 pp, Deep Neural Networks by 2.76 pp and small Convolutional Neural Networks by 7.28 pp. A Friedman omnibus test rejects equal rank with chi^2 = 23.71 and p = 2.46 x 10^-4. Paired t-tests reach p < 10^-2 against three of the five baselines after Bonferroni correction (MLP, DNN, CNN), with Random Forest borderline. The contribution is biological and empirical. We re-anchor GF to its RNA roots. We provide the first GF-family algorithm to reach the top mean rank against five classical baselines on a vision benchmark.