By its nature, deep learning can not generalise well from a few examples. This is why the detection of rare objects is a difficult problem. As industrial demand for inexpensive learning on limited data grows, many new sophisticated few-shot learning methods appeared, albeit few, if any, can challenge a simple transfer learning strategy.
In order to present and validate results from this murky area of machine learning, a state-of-the-art method called the Two-stage Fine-tuning Approach is implemented and tested. Moreover, two improvement areas - model fine-tuning adaptation and data augmentation - are identified and shown to even double the performance compared to the baseline on low-shot tasks.
Dissertation required for completion of M.S.c degree in Artificial Intelligence at University of Limerick.
Find more here.