Contrastive learning for dimensionality reduction and visualization of transcriptomic data

01/04/2021 to 01/04/2025

While most successful applications of machine learning to date have been in the realm of supervised learning, unsupervised learning is often seen as a more challenging, but possibly more important problem. Yann LeCun famously compared supervised learning with the thin icing on the "cake" of unsupervised learning. An approach called contrastive learning has recently emerged as a powerful method of unsupervised learning of image data, allowing, for example, to separate photos of cats from photos of dogs without using any labeled data for training. The key idea is that a neural network is trained to keep each image as close as possible to its slightly distorted copy and as far as possible from all other images. The balance between attractive and repulsive forces brings similar images together. In this project we will apply these ideas to single-cell transcriptomics, a very active field of biology where one experiment can measure gene activity of thousand of genes in millions of individual cells. We will use contrastive learning to find structure in such datasets and to visualize them in two dimensions. We will then go back to the image data and use two-dimensional embeddings as a tool to gain intuition about how different modeling and optimization choices affect the final representation.