2023_programme: Few-Shot Learning for ATR using Contrastive Loss and Modeled Data: Understanding Fidelity Requirements

Back


  • Session: 08. Munition Detection- Classification and Localization in the Marine Environment
    Organiser(s): David Bradley
  • Lecture: Few-Shot Learning for ATR using Contrastive Loss and Modeled Data: Understanding Fidelity Requirements
    Paper ID: 2092
    Author(s): McMillan Justin
    Presenter: McMillan Justin
    Abstract: One of the largest drawbacks of Deep Neural Network (DNN) implementation is the amount of data required to train the network. Building sufficiently large datasets for training can not only be expensive and time-consuming, in some cases it is simply not possible. To circumvent this obstacle, various few-shot learning techniques have been developed, with the most successful methods incorporating synthetic data to augment scarce training samples. In this presentation, we compare different approaches to few-shot learning, and quantify how simulation fidelity affects classification performance on “real” samples. We find that not only does Domain Adaptation outperform other techniques significantly, we also present evidence suggesting that Domain Adaptation is much more robust to fidelity degradation.
  • Corresponding author: Mr Justin McMillan
    Affiliation: Applied Research in Acoustics
    Country: United States
    e-mail:

Back