We used a linear autoencoder (LAE) and its learning dynamics to analyze the high-order structure of human mitochondrial DNA (mtDNA). A total of 360 complete human mtDNA sequences were collected from the MITOMAP database and transformed into 1024-dimensional vectors of pentanucleotide frequencies. We compressed those into a three-dimensional (3D) coordinates by an LAE at each step of training by gradient descent with respect to the quadratic error function. Along the time axis of training epochs, the compressed 3D coordinates were gradually clustered and separated in accordance with the order of the genetic distance in the phylogenetic tree of human mtDNA haplogroups. This suggests that there is an association between the learning dynamics of LAE and the high-dimensional structure of human mtDNA sequences, similar to that of phylogenetic analysis and evolutionary pathways: the five clusters eventually contained only a single haplogroup of L0, M, N, R, and U, while the L3 cluster contained a small number of M members and The packing was comparable to that realized in learning dynamics similar to genetic classification and evolutionary pathways by LAE in principal component analysis (PCA), but somewhat denser than PCA.