Weakly Supervised Multi-Task Representation Learning for Human Activity Analysis Using Wearables
summary
Proposes a weakly supervised multi-output siamese network that learns separate but linked embedding spaces so one space groups activity types and another groups person identities, and adds a gender embedding in a later experiment. Trains with pairwise similarity signals and even with halves of the data providing only activity or only identity relations (partial similarity supervision), finds the multi-task model often outperforms single-task supervised baselines, scales to extra attributes, and ablations show no performance loss and frequent gains.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 0:39Limitations
- 2:44Representation Spaces
- 3:08Results
- 3:29Second Experiment
- 4:01Third Experiment
- 5:09Conclusion
related talks
06mUnsupervised Human Activity Representation Learning with Multi-task Deep Clustering
28mDeep Learning-Based Human Activity Recognition for Physical Skills...
06mMultimodal Joint Head Orientation Estimation in Interacting Groups via Proxemics and Interaction ...
06mSelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data
10mLearning Network-Based Multi-Modal Mobile User Interface Embeddings
06mContrastive Predictive Coding for Human Activity Recognition
05mLeveraging Activity Recognition to Enable Protective Behavior Detection in Continuous Data
21mGraph Representation Learning with Deep Embedding Approach