FROM ZERO TO DEEP LEARNING WITH SCALA
summary
Raspberry Pi camera collects ~14,000 labeled images cropped to 171x25 pixels; model built in Scala with DeepLearning4J and pipeline implemented with Akka Streams. Network uses 5x5 conv filters (20 then 50 feature maps), two subsampling layers, a 500‑unit dense layer and 2‑way output. Training runs with early stopping (~6 minutes, 8 epochs) and best score 0.964 at epoch 5. Precision stays high but recall is low and variable; common failures come from weather, partial planes and a bias toward blue (KLM) aircraft.
AI-written summary · about
chapters (17)
- 0:00Intro
- 3:09View From My Office
- 3:55Raspberry PI setup
- 9:34Problems Matched By Architectures
- 11:52The Artificial Neuron
- 13:45Multilayer Neural Network
- 18:07The Power Of CNNS
- 20:20How Does A Convolutional Layer Work?
- 22:13A Typical CNN Architecture (LeNet)
- 25:06Training Images - Crop And Resize To 171 x 25 Pixels
- 26:59Some settings
- 27:42Our Network In Scala Layer By Layer
- 31:58Train And Evaluate Our Network
- 33:59Logs of a training run
- 35:58Al is only a small bit of Al projects
- 43:03When does it go wrong?
- 48:05Model Performance
related talks
39mDive into Deep Learning with Scala
18mDeep learning for Space imagery applications
34mDeep Learning at Scale
37mUnstructured Data Processing with a Raspberry Pi AI Kit and Python
36m"Lights, camera, AI!" - Automated sports videography
48m(Deep) Learning to Fly
44mAdventures in Puppy Training with A.I. and a Raspberry Pi
54mDeep Learning with Apache Spark