(A first step toward) mechanistic interpretability through the training of a small-scale sparse autoencoder (144 photos at 64x64px) to identify features within the set of images and mapping neurons to interpretable concepts. Includes a set of visual representations for each feature by identifying images that have the highest activation values.