Self-supervised Embeddings

To accelerate analysis and exploration of H01, we produced embeddings for local cutouts with a neural network. This network was trained using a variant of the SimCLR self-supervised learning framework, meaning no labels were presented during training. We computed embeddings for 20% of all skeleton nodes (~4 billion in total) and reduced them to 3 dimensions using UMAP for visualization and color mapping. This interactive scatter plot shows a sample of these embeddings. Clicking on a point brings up a neuroglancer view showing the associated location in H01. The c3 embeddings can also be explored in neuroglancer directly.