Contrastive Predictive Coding for Human Activity Recognition
summary
Applies Contrastive Predictive Coding to wearable-sensor HAR using a 1D convolutional encoder and GRU to produce z-vectors and a context Ct, trained with InfoNCE loss to predict k future timesteps. Benchmarks on MobiAct, MotionSense, UCI-HAR and USC-HAD show CPC outperforms supervised DeepConvLSTM on two datasets and beats unsupervised baselines on three, and it yields roughly +10% F1 when only 10 labeled windows per class are available.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:15SENSOR DATA ANNOTATION IS MORE DIFFICULT THAN DATA COLLECTION
- 0:52LEVERAGING UNLABELED DATA FOR LEARNING
- 1:34OVERALL SELF-SUPERVISED LEARNING WORKFLOW
- 2:04STAGE I: CONTRASTIVE PREDICTIVE CODING
- 3:01STAGE 2: ACTIVITY RECOGNITION
- 3:13DATASETS
- 3:24ACTIVITY RECOGNITION PERFORMANCE
- 4:32SEMI-SUPERVISED LEARNING WITH LIMITED LABELED DATA
- 5:44CONTRASTIVE PREDICTIVE CODING FOR HUMAN ACTIVITY RECOGNITION
related talks
06mA Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition
52mPredictive Coding Models of Perception
28mDeep Learning-Based Human Activity Recognition for Physical Skills...
06mWeakly Supervised Multi-Task Representation Learning for Human Activity Analysis Using Wearables
06mCross-Dataset Activity Recognition via Adaptive Spatial-Temporal Transfer Learning
06mAdversarial Multi-view Networks for Activity Recognition
06mRobust Unsupervised Factory Activity Recognition with Body-worn Accelerometer Using Temporal ...
05mMETIER: A Deep Multi-Task Learning Based Activity and User Recognition Model Using Wearable Sensors