Multiple instance learning is a set of problems in which types of objects can be represented as "instances of bags" (i.e., types of apples, pears, peaches as instances of apples, pears, and peaches as bags of fruit). One common problem involving multiple instance learning is in drug development. Drug molecules can take on different shapes called classifications which have different activation properties. These classifications can be represented as instances of the molecule (the bag) with labels representing the degree of activation.
This project is my research project with Dr. Soumya Ray at Case Western Reserve University. Dr. Ray and his students previously developed a unique algorithm for approaching the multiple instance learning problem called MIRK (multiple instance regression with kernels). This algorithm proved to be more efficient than every other existing method for this problem... except for single instance learning. I'm exploring what happens during training that causes the simple single instance learning algorithm to outperform the MIRK algorithm.