TLDR: Just created a citiziens science website where you can help contribute to research. Please try it out. There's a short training and then a review tab where you can actulaly contribute to real event detection.
https://lenswatch.dev
Here's the longer but more in depth description: We trained a model on real OGLE survey data to flag microlensing candidates which are astronomical events that can lead to dark matter, black hole, and other cool object detection findings. It's confident on the obvious ones, but there's a messy middle where it genuinely can't tell. Those are the ones we send to people.
You look at a light curve (a graph of a star's brightness over time) and answer a few yes/no questions about the shape. The site walks you through what to look for, so you don't need any prior knowledge. Takes about five minutes to learn.
When volunteers disagree with each other on a curve, we don't throw that out. The disagreement itself becomes a signal the model learns from, so it gets better at knowing which cases are genuinely ambiguous instead of just guessing confidently and being wrong.
TLDR: Just created a citiziens science website where you can help contribute to research. Please try it out. There's a short training and then a review tab where you can actulaly contribute to real event detection. https://lenswatch.dev
Here's the longer but more in depth description: We trained a model on real OGLE survey data to flag microlensing candidates which are astronomical events that can lead to dark matter, black hole, and other cool object detection findings. It's confident on the obvious ones, but there's a messy middle where it genuinely can't tell. Those are the ones we send to people.
You look at a light curve (a graph of a star's brightness over time) and answer a few yes/no questions about the shape. The site walks you through what to look for, so you don't need any prior knowledge. Takes about five minutes to learn.
When volunteers disagree with each other on a curve, we don't throw that out. The disagreement itself becomes a signal the model learns from, so it gets better at knowing which cases are genuinely ambiguous instead of just guessing confidently and being wrong.