Lessons learned using Machine Learning in Java
summary
Builds a Lego block visual sorter using Java libraries (Deep Java Library, MXNet, TensorFlow engine) and transfer learning, trains on ~1,800 augmented images per class, and targets the top 1,000 bricks. Benchmarks show Java outperforms Python slightly for training and inference; training 100 classes takes ~9 hours for 25 epochs. Zoomed‑in rendered plus a small set of real photos yields 80%+ accuracy for individual bricks and a conveyor prototype that reliably sorts with tuned confidence thresholds.
AI-written summary · about
chapters (19)
- 0:00Introduction and motivation
- 1:02Choosing Java over Python
- 3:10Java ML frameworks overview
- 6:24Dog vs muffin demo
- 8:10Data set preparation challenges
- 12:27Building a model in Java
- 17:47Training and evaluation metrics
- 21:23Saving and using the trained model
- 25:41Early stopping and model tuning
- 29:14Optimizers and transfer learning
- 31:24Using Python models in Java
- 34:17Lego sorting project setup
- 37:07Generating Lego brick data set
- 38:30Model accuracy with zoomed images
- 40:34Real images vs generated data
- 41:58Grayscale conversion and contrast
- 43:48Data augmentation techniques
- 45:56Final model performance
- 47:04Key takeaways and conclusion
related talks
43mLessons Learned Using Machine Learning in Java
37mJavaScript Machine Learning from Magic to Science
58mAn introduction to Machine Learning using LEGO
58mAn introduction to Machine Learning using LEGO
46mML in Java, YES it's possible!
50mML in Java, YES it's possible!
19mDeep Learning is Just LEGO: & Other Hands on Machine Learning Activities
47mML in Java, YES it's possible!