When teaching a student, you sometimes have to come up with more than one way of explaining the same concept. The same ideas apply when teaching a machine how to interpret images. We have images from a rock sample, each of which contains pixels that depict the fossil remains of an extinct reef building organism called archaeocyathids. The 3D structure of the 515 million year old reef formed by these archaeocyathids provides valuable information, not only on the ancient reefs, but also about the impact of modern reefs on Earth’s future biosphere. Recovering this 3D structure involves identifying the archaeocyathid pixels in each image, a time-consuming process. Additionally, the machine learning model that we apply to automate this process does not fully understand what an archaeocyathid looks like. Thus, we introduce two methods to assist the model’s learning process and find that providing the model with these alternate ways of thinking about the images in our dataset improves its performance, bringing us one step closer to understanding the impact of reefs.