Research Project Title:
Augmented One-Shot Learning
abstract:There is a push to make beyond-human capable robots that can do things faster and to greater precision than humans can, but there is a huge disparity between how humans perceive the world and how robots do. In our project we strive to give robots the ability to see novel objects and be able to recognize similar objects in the future the same way a human is able to. We are combining traditional one-shot matching networks (such as Mataching Networks, Prototypical networks, etc.) with prior knowledge obtained via a Hierarchical Nonparametric Bayesian Model, to improve the accuracy of any one-shot solution existing today.
"My previous UROP experience has been in traditional perception algorithms and determining how much data is needed to add additional classes to pre-trained models. This has lead me and my advisors to find new techniques that require less data and hope to learn how language helps humans learn so much faster than machines. I am excited to push perception as far as we can and strive to achieve human capabilities in vision."