Mobiconf 2018: Miłosz Kaszyński, OKE Poland
summary
Claim: phones run inference not training, so use them as data collectors and private consumers of models, not model builders. Practical tools: convert models with Core ML / TensorFlow Lite after freeze and optimize; expect many conversion failures. Favored models: MobileNet SqueezeNet ShuffleNet and small LSTMs; avoid >50-layer ResNets. Measured tradeoffs: Volta-class GPUs hit ~6,000 fps, same model on iPhone X takes ~120 ms; phones handle one image at a time, battery and memory limit throughput. Example: OpenPose + LSTM fall detector trains in hours and runs ≈6 fps on embedded GPU; works but slow.
AI-written summary · about
chapters (9)
- 0:00Introduction and background
- 2:14Why machine learning on mobile
- 8:23Mobile ML model limitations
- 17:51Performance and hardware constraints
- 23:50Device compatibility and frameworks
- 30:26Challenges in model conversion
- 33:42When to use mobile ML today
- 36:19Future of mobile AI technology
- 39:22Q&A session
related talks
27mMachine Learning and Mobile
39mCreating smaller, faster, production-worthy mobile machine learning models for Android
14mOptimize your TensorFlow Lite models
10mMachine Learning for mobile
32mAI for Mobile and IoT Devices: TensorFlow Lite
15mSimplified Machine Learning on mobile, browsers, and edge devices with on-device ML
40mDeep Learning on iOS
08mPyTorch Mobile