Unsupervised Human Activity Representation Learning with Multi-task Deep Clustering
summary
An end-to-end pipeline compresses raw wearable sensor windows with a CNN-BiLSTM autoencoder, assigns K-means pseudo-labeling on the latent space, and trains a two-layer DNN classifier with a joint multi-task loss that optimizes reconstruction, clustering distance, and classification simultaneously. Evaluation on three public datasets reports up to 30% improvement over shallow unsupervised baselines, near-supervised accuracy after small fine-tuning, and reasonably stable cross-dataset transfer when trained on MobiAct.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:17Research Problem
- 0:52Methods
- 1:42Overview
- 2:17Autoencoder Structure
- 2:39Kmeans Deep clustering
- 3:06Deep DNN classifier
- 3:32Endtoend model
- 3:48Experiment
- 4:07Results
- 4:50Visualization
related talks
06mWeakly Supervised Multi-Task Representation Learning for Human Activity Analysis Using Wearables
28mDeep Learning-Based Human Activity Recognition for Physical Skills...
06mAdversarial Multi-view Networks for Activity Recognition
06mA Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition
05mMETIER: A Deep Multi-Task Learning Based Activity and User Recognition Model Using Wearable Sensors
28mEnhancing Upper Limb Rehabilitation with Unsupervised Machine Learning Models
06mCross-Dataset Activity Recognition via Adaptive Spatial-Temporal Transfer Learning
34mComputer Vision and NLP for Multi-Task Fashion Attrb. Modeling