Salmon Counting · Episode 4 of 8 · 37 min
How Humans Teach AI to Count Salmon on Video
How does a machine learn to see a salmon? One click at a time. In a University of Washington lab on Lake Aleknagik, a technician clicks each time a fish crosses a virtual count line on drone video, and thousands of labels teach a model to find, track and count salmon. The episode explains bounding boxes, training rounds, glare, shadows and murky water, and why a human stays in the loop.
Who's who
- Jessie Sheldon+1 more episodeReporter whose July 2026 story is the episode's source
Where it happens
- Lake Aleknagik, Alaska+2 more episodesField lab where the labeling is done
- Bristol Bay, Alaska+7 more episodesRivers counted for more than 70 years from towers
Timeline
- 2019British Columbia researchers collect the data behind the SalmonVision dataset.
- October-November 2023A deep-learning salmon detection and tracking paper is published.
- Summer 2026Labeling and model training proceed at the Lake Aleknagik lab.
- July 8, 2026Alaska Public Media publishes the report on AI, drones and salmon.
Still debated
Accuracy against the tower counts is still pending, and no accuracy figure for this project exists yet. Whether a model trained under one set of conditions generalizes across rivers and conditions is also a concern.