AI is for the birds
summary
A DIY system uses MobileNetV2 with transfer learning to classify squirrel vs not-squirrel, trained on about 140 images drawn from ImageNet, yielding roughly 90% accuracy on held-out sets. Training runs on a laptop CPU; inference can be offloaded to a USB TPU. Detection triggers an audible deterrent that proves ineffective in the field; no live-squirrel images were captured for retraining and hardware issues (Raspberry Pi camera, TensorFlow vs PyTorch) complicate deployment.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:24Squirrels at bird feeders
- 3:53Squirrel Away project
- 4:57Convolutional neural networks
- 9:17Transfer learning MobileNet V2
- 12:56Training AI for squirrel detection
- 15:09Live demo training
- 17:01How inferencing works
- 18:07Hardware TPU setup
- 21:01Live demo detection deterrent
- 23:22Lessons learned
- 24:25Mechanical squirrel behavior
- 25:07Training data sources
- 27:32Dataset size splitting
- 28:59Project future plans
- 30:05Hotel room electrical mishap
- 31:53Retraining with real images
- 33:58Image dataset size rule
related talks
41mAI in Web: TensorFlow js
23mBirding with Python and Machine Learning
31mBuild an AI-powered Pet Detector in Visual Studio Code
10mDeep Learning in the Wild: Embedded PyTorch for...
01h 02mAI on the Edge
23mAI is ready for you. Are you ready for AI?
38mReal-world Machine Learning with TensorFlow and Cloud AI
22mAI On The Farthest Edge