Dmitry Gavrilov Qualofying Exam (QE)

Abstract: Leveraging Dynamics of Feature Space for Effective Semi-supervised Learning with Scarce Data

This study introduces a novel semi-supervised learning approach focused on analyzing feature
space transformations of a classifier during training and on examining how feature space evolution
influences classifier development. The idea is to treat embeddings as moving through the
feature space, using their displacement vectors to infer class identities. This concept of “drift
of embeddings” allows automatic labeling of new data samples, reducing the need for typical
human labeling in supervised and semi-supervised learning paradigms. The preliminary results,
including tests on the CIFAR-10 dataset, show the viability and superiority of the method over
the traditional k Nearest Neighbours algorithm. In particular, this method holds promise for
creating datasets in agriculture, especially for drone-based crop recognition, where manual labeling
is often challenging, time-consuming, and costly. The study identifies method limitations
and suggests future research directions.