Modeling Many-to-Many Maps

Machine learning models have obtained impressive performance on benchmark many-to-one datasets such as CIFAR-10. However, there are many real-world datasets that can be described as many-to-many. That is, a single input can yield different outputs and many different inputs can yield the same output. For example, in an image-to-text task, a single image has many possible captions and conversely many different images can have the same caption. Building off of previous work that used fiber bundles to model many-to-one processes, we present a model for many-to-many processes and discuss some of the challenges of working with real-world data that motivated this work.

This talk will be hybrid, held in-person and online on Zoom