Cross-Image Context Matters for Bongard Problems

Thanks for the post! I'm the researcher and would be happy to answer questions if people have them!

Here's a brief summary: Bongard problems are a type of visual reasoning task where the human/AI sees two sets of images: a “positive” set where each image depicts a certain concept and a “negative” set where images don’t depict the concept. The goal is to determine the concept given these sets. Unfortunately, while humans can often solve these problems in the ~90% accuracy range, AI performance is often in the ~60% range.

In our work, we devise simple methods that incorporate "cross-image context" to attain higher performance. "Cross-image context" means that we consider the similarities and differences between multiple positive and negative images. We find that performance of simple existing methods (e.g., k-nearest neighbors on deep features) can be dramatically improved by incorporating cross-image context, sometimes up to 10%!

In all, our work attains state-of-the-art performance on the two main Bongard datasets.

Let us know what you think!