Deep Dive into Deep Learning with KotlinDL
summary
Live demo of KotlinDL shows actual code: ResNet50 transfer learning where the top layers are replaced, trained on a tiny dataset, and accuracy rises from ~30% with the empty head to usable levels within 50 cats/50 dogs, 8 epochs (plots suggest 2–4 epochs often suffice). The talk uses the JVM TensorFlow backend, model hub, built-in preprocessing, saves model JSON/weights, and highlights ONNX export and Android inference (no on-device training). Concrete demos, limits, and export formats are shown so it's clear what can be reused.
AI-written summary · about
chapters (12)
- 0:00Introduction to KotlinDL
- 2:24Challenges in ML with Kotlin
- 5:08KotlinDL library features
- 7:18How KotlinDL works
- 9:57Object detection demo
- 15:34Pose estimation demo
- 17:55Face landmark detection demo
- 19:33Image recognition demo
- 23:03Transfer learning concepts
- 25:15Transfer learning demo
- 42:48Saving and loading models
- 47:20Summary and resources
related talks
58mTwo Years in the Life of a Deep Learning Library in Kotlin
47mDeep Learning: a practical introduction from an Android dev perspective
44mKotlin strives for Deep Learning
01h 04mWhat you always wanted to know about Deep Learning, but were afraid to ask
39mMachine Learning on Android Demystified
46mFire Up Your Smartphone Neurons with On-Device Deep Learning
42mCreating Smarter Applications
35mDeconstructing Deep Learning