Enhancing OrcaAL Active Learning Tool

[valentina-s]

During GSOC 2021 @jd-rs and @kunakl07 developed an active learning system called OrcaAL which visualizes predictions of a machine learning algorithm on new data and allows input from the user to correct and annotate more data, which then supplement the existing machine learning training dataset.

Repos:
https://github.com/orcasound/orcaal
https://github.com/orcasound/orcaal-research
Blog Posts:
https://www.orcasound.net/tag/gsoc/

There are several directions one can improve the system:

• Add new sampling strategies to Orcaal, and analyze performance
• Currently the sampling strategy is based on selecting samples on which the algorithm is most uncertain
• Add Tensorboard visualizations to Orcaal for ML experts
• So far we have performance plots aimed at a general audience, but ML experts maybe interested in more detailed performance plots (those can show in a separate window)
• Tensorboard also allows audio snippets to be shown:
• Perform more analysis on the annotation data:
• The annotation data is stored in a database run by OrcaAL
  • psql postgresql://postgres:postgres@localhost:5432/orcagsoc
• Allow to change source of data
• Experiment with the full Orcasound stream, or OOI stream
• Allow to change model
• Experiment with dockerized models:
  • How can we log the parameters
  • Can we easily add other models such as for ship, click detection, etc.
  • Improve performance:
    • GPU processing in a separate instance
    • How can we spin up a separate instance using cloud resources
• Test performance when many people annotate simultaneously

Possible Mentors: Valentina @valentina-s, Kunal @kunakl07

Getting Started:

• Install the app locally
• Create a new docker image with a modification of the model, interface, etc
• Rerun the app with the new docker image