One of the fundamental difficulties in contemporary supervised learning approaches is the dependency on labelled examples. Most state-of-the-art deep architectures, in particular, tend to perform poorly in the absence of large-scale annotated training sets. In many practical problems, however, it is not feasible to construct sufficiently large training sets, especially in problems involving sensitive information or consisting of a large set of fine-grained classes. One of the main topics in machine learning research that aims to address such limitations is few-shot learning where only few labeled samples are made available for each novel class of interest.
An inherent difficulty in few-shot learning is the various ambiguities resulting from having only few training samples per class. To tackle this fundamental challenge in few-shot learning, in this thesis, we propose an approach that aims to guide the meta-learner via semantic priors. To this end, we build meta-learning models that can benefit from prior knowledge based semantic representations of classes of interest when synthesizing target classifiers. We propose semantically-conditioned feature attention and sample attention mechanisms that estimate and utilize the importance of representation dimensions and training instances. In sample attention, we aim to weigh each individual training example based on its representativeness for the related class. We, then, use the information extracted from each example proportional to its individual weight. In feature attention, we aim to weigh each visual feature dimension based on the semantic embedding vectors we obtain for each class. We also study the problem of sample noise in few-shot learning, where some training examples are irrelevant due to annotation or data collection errors, which can be the case for various real-world problems.
Our experimental results demonstrate the effectiveness of the proposed semantic few-shot learning model with and without sample noise.