Ted Willke, Sr. Principal Engineer, Intel
summary
Uses low-flying snotbot drones and shore photos to do two practical ML systems on Alaska 2017 expeditions: a fluke-ID pipeline that extracts trailing-edge geometry, runs a lightweight classifier on a laptop over wireless to return top-10 matches (found a whale not seen for 21 years), and a morphometry pipeline that adapts an ImageNet backbone with a few hundred whale frames plus one-shot transfer learning to segment bodies and estimate energy stores; works in messy water but segmentation error still amplifies volumetric estimates.
AI-written summary · about
chapters (22)
- 0:00Intro
- 1:13Humpback whales
- 2:26Snot bot
- 3:43Whales
- 5:32Humpback
- 6:06Right Whales
- 8:09Whale Mysteries
- 9:33Missing Data
- 10:35Old School
- 11:28Ocean Alliance
- 12:47Collisions
- 13:17Whale Identification
- 14:02Morphometry
- 15:41Real World Example
- 16:22Key Challenges
- 17:38Data starvation
- 18:33Photo ID
- 18:52Avoid Deep Learning
- 20:21Early Results
- 20:44Top Rankings
- 26:12Example
- 26:32Call to Action
related talks
24mAI in maritime - Autonomous ships
41mKEYNOTE: Image processing, AI, and autonomous systems
25mDeep learning in deep nets: Helping fish farmers feed the world
40mExploring deep space and deep ocean with AI guidance
36mDeep Learning - Segmentation, Identification and Classification
31mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet -An Open Source Workflow
01h 09mFrom Walruses to Code: A Guide to Tackling Unique ML Problems
32mThe ML system endeavors