Muvr
summary
Code runs on wearables and a mobile app does signal processing sensor-data fusion and classification; models and CSV data land on a server where Spark-based training on EC2 cluster fits multi‑layer perceptrons and writes weights back to S3. Real constraints: 1200 inputs (400 per axis), sliding 2s windows stepped every 50ms, on‑phone forward propagation to avoid ~500ms latency, ND4J/BLAS for fast vector ops. Collected dataset and code are published; surprising result is 60% of users supply unusable data, so pipeline emphasizes curation and personalized models.
AI-written summary · about
chapters (27)
- 0:00Introduction
- 0:21What is Muvr
- 1:15Recording accelerometer data
- 2:01What I thought
- 2:19How we went about it
- 2:45Training data
- 2:56Results
- 3:09Classifier
- 3:19Exercise
- 4:26Testing
- 7:14Whats missing
- 8:44Perceptron
- 9:23Network
- 9:51Inputs
- 11:10Sliding window
- 11:54Classification
- 12:34Demo
- 18:34Optimization
- 20:54Pipeline
- 22:37Preprocessing
- 23:21Spark Context
- 24:27Model Templates
- 25:32Initial Model
- 26:23Train Model
- 39:43Restarting
- 41:20Project Overview
- 48:57User experience
related talks
44mExercise Analysis
42mSo, You Want to "Do Big Data?"
17mAI workflow for large scale deployment of far-edge ML devices
26mNext generation machine learning for mobile and embedded platforms
33mAnalytics in the age of the Internet of Things #bbuzz
41mBeefy ML: Ultra-Low Power Algorithms on Cattle
56mMAAS: Mind as a Service
41mBeefy ML: Ultra-Low Power Algorithms on Cattle