Multiple Instance Learning #23
Description
https://arxiv.org/pdf/2506.09022
This paper utilizes multiple instance learning. The idea is that you train a bunch of weakly supervised classifiers on small amounts of data.
This probably means it would be the train data for benchmarks.
The benefit of multiple instance learning is that it allows us to use much smaller amounts of data, the data is focused on the task in question, and the supervision allows us to expect greater performance.
We want to consider various MIL choices.
This issue is left ambiguous - if you are interested, come up with a method, and discuss in the channel.