Abstract
<jats:p>In this paper, we demonstrate that it is feasible to train a deep generative model for dimensionality reduction with millions of features using relatively few samples, making this type of generative model a more versatile alternative to standard dimensionality reduction methods. Specifically, we hypothesize that, for a decoder-only model, the number of training samples required is almost independent of feature dimensionality across most network architectures. Through an extensive set of experiments on synthetic nonlinear data, we validate this hypothesis. We further train the model on a downsampled version of the 1000 Genomes Project (1KGP) dataset to assess its behavior under controlled reductions in sample size. Finally, we train a Deep Generative Decoder (DGD) on a curated dataset from the International Cancer Genome Consortium (ICGC) containing 4.4 million features, using approximately 4,000 training samples and 1,000 test samples. The resulting latent representation exhibits clear clustering and, when compared at the same latent dimensionality, outperforms PCA and variational autoencoders (VAEs) for tumor type classification. In addition, the DGD is computationally efficient and can be trained on a single 16 GB GPU. The implementation is available at https://github.com/cpancott/ReceptiveDGD.</jats:p>