A Feasibility of On-Device ML for Adaptive User Interfaces in Frontend Development
summary
Claims feasibility of real-time, client-side personalization using small models (trained ~50 epochs on ~10–12 synthetic sessions) that load in under a megabyte and run with millisecond inference via JS, WASM, WebGL or WebGPU backend. Demonstrates a dashboard demo that predicts next-clicks and reshuffles cards (causing a “creepy” UX until an explicit indicator and manual mode are added). Notes toolchain options like TensorFlow.js and ONNX runtimes, records clicks/hover time and time-of-day features, reports 80% of users never change defaults, and recommends privacy by design, transparency, and opt-out/reset.
AI-written summary · about
chapters (13)
- 0:00On-device ML for adaptive UIs
- 3:13Preference fatigue in UIs
- 7:28Adaptive UIs that learn
- 9:38Why choose on-device machine learning
- 11:48On-device ML tools in browsers
- 13:35Demo: adaptive dashboard with TensorFlow
- 18:14Improving adaptive dashboard
- 20:23How adaptive UI works
- 21:29ML performance backends
- 22:53Efficient data for adaptive UIs
- 23:16UX and ethics for adaptive UIs
- 25:26When to avoid adaptive UIs
- 27:10Conclusion
related talks
19mML at Client Side
15mThe future of machine learning using JavaScript
29mSuperpowers for next gen web apps: Machine learning
41mPractical Machine Learning for Front-End Developers
48mDeep Learning in the Browser
25mUsing Machine Learning To Improve UI/UX
41mEscaping the Cloud: High-Performance AI in your Browser
39mMind-Reading with Adaptive and Intelligent User Interfaces