Applied machine learning: a few lessons I learned the hard way
summary
Machine learning often fails from data issues not model complexity: a nearest neighbor classifier can be written in two lines and solve simple tasks, but overfitting ruins results without representative data. Concrete fixes: use data augmentation to turn 1,000 images into ~100,000 so validation tracks training, convert to grayscale to fix a color-space mismatch in a tunnel installation, and expect ~85% accuracy on a drone trail classifier (matching human performance). Tools and libraries mentioned include Keras and imgaug.
AI-written summary · about
chapters (9)
- 0:00Start
- 1:46The Classification Problem informal definition
- 6:57Step 2: inference
- 29:29An image classification problem
- 29:57Classification accuracy
- 33:31Implementation
- 36:19The problem
- 37:58Training data augmentation
- 40:47Conclusions
