Module 40: Artificial Intelligence for Autonomous Interstellar Spacecraft Novelty Detection

Faculty Contact: Linda Petzold

Research Areas:

Abstract (Phase 1): This project essentially lives at the crossroads of Experimental Cosmology & Applied Computer Vision. Most of the actual work will be hypothetical for now, but in general, we want to send out a series of probes that will venture deep into space. For perspective, they probably wouldn't even turn on until they have traveled beyond our solar system. While traveling, they will encounter new & interesting things in space, such as unseen planets, asteroids, suns, or moons. The first part of the project relies on computer vision in order to spot these astronomical objects and recognize that they are of a certain type. A never-before-seen planet should be recognized as a planet, despite never being trained specifically with images of that planet. Since all of this hinges on the probe discoveries being vastly new and unseen, this delves into a budding area of computer vision called 'Unseen Object Recognition'.

Abstract (Phase 2): Assuming we can get a good object detection algorithm here, the second hurdle is defining an 'Interesting Metric'. We pose the question as: "You are shown two images, both containing different, unique, and never-before-seen planets. We can only keep one of them. Which one is more interesting to you?" Human beings wouldn't fully agree here, and if put into an actual experiment, different humans will definitely choose different planets. Why is that? Currently, we are working with Brain Psychologists at UCSB to set up a possible experiment to gain qualitative insights and data behind what the brain goes through in order to classify what humans think is Interesting (in a space context for this project). The hope is that once we understanding some of the workings of the brain, and why it makes decisions like this, we can then train a machine to look out for the same clues and deduce an Interesting Score, most likely from image features and human-based weighted functions.

Active Quarters:

  • Winter 2018: James Bird
  • Fall 2018: James Bird