Continual learning for Human Activity Recognition

Human Activity Recognition (HAR) has been developed for a long time. Many highly accurate models were proposed. However, one major obstacle towards the real-life smart home is the ability to adapt to change.

This project aims to develop a continual learning model that is able to learn a new task but also retain previously learned tasks. The key technique used here is generative replay. Two agents are working together: one generative model is to mimic all trained lesson while one fully-connected neural network is to gradually lean to capture structured knowledge.

See technical detail here

It is an part of CS5099 Dissertation at University of St. Andrews.

Set up

1. Prepare python environment python3 -m venv env
2. Enter to the env source env/bin/activate
3. Install library pip install -r requirements.txt

Run experiments

Run python run_main.py --results-dir="YOUR RESULT DIR" --data-dir="DATASET"

DATASET must be one of these options ["casas", "housea", "pamap", "dsads"]

Other available arguments can be seen by running python run_main.py --help

You could run run_*.py for a different experiment (The same arguments can be applyied in any experiments). You need to run visdom if you want to run run_plot.py.

Visdom

1. Run python3 -m visdom.server in one terminal
2. Open browser and then enter url given by the visdom server
3. Open anohter terminal then run your model with --visdom (only work with run_plot.py)
4. See the interactive report

• implement XdG
• consider diversity as well as accuracy
  • self-verifying => select samples with probability rather than yes/no result
• GAN-ralated
  • implement Minibatch or Feature mathcing to avoid mode collapse
  • try Unrolled loss https://arxiv.org/abs/1611.02163
  • label smoothing https://arxiv.org/abs/1606.03498
  • TTUR Two Time-Scale Update Rule https://arxiv.org/abs/1706.08500
    • different learning rates for the generator and discriminator
    • 0.0004 for the discriminator and 0.0001 for the generator
• automated parameter selection
  • fANOVA
• try other GANs
• try hierarchy model
• experiments with task evolution
• experiments with noise; robustness to noise

Author

Pakawat Nakwijit; An ordinary programmer who would like to share and challange himself. It is a part of my 2018 tasks to open source every projects in my old treasure chest with some good documentation.

License

This project is licensed under the terms of the MIT license.