k-Shot Learning for Action Recognition

In the problem of k-shot learning, a model must learn to reliably classify an example having seen only k previous instances of examples of the same class. With recent success in using memory in neural networks to perform kshot learning, we propose a technique that uses MemoryAugmented Neural Networks to perform k-shot learning for action recognition in videos. We believe the use of memory will help learn encoded generalizable representations of actions that can be used for classification. We use the Memory-Augmented Neural Network framework to evaluate k-shot learning for the Kinetics Dataset of action clips taken from YouTube videos. Our approach works well for this domain, achieving results close to 78% accuracy in the 9-shot learning case.