The nanometer-resolution, 3D image volumes of brain tissue reconstructed using high-throughput electron microscopy (EM) are beginning to generate important insights about how neural circuitry is built, and about how that structure relates to neural function. This technique is unique in that it simultaneously resolves all cells in a tissue, their organelles, and their connections (chemical synapses), which makes it possible to build a very detailed picture of patterns in neuronal connectivity. However, even in small EM image volumes there can be hundreds to millions of any single type of biological object, so that it is impossible for humans to assess all examples and exhaustively define and measure their important features. In this talk, I will discuss an effort to circumvent human-based feature design by generating unsupervised feature representations of adult chemical synapses in the brain (in an EM connectomics reconstruction of mouse primary visual cortex) using contrastive deep learning. In particular, I will discuss several salient patterns in synapse features we observe that support this approach as a scalable way to make sense of neural connectivity as connectomics datasets become larger.