Paper - Towards Environment Independent Device Free Human Activity Recognition

  • Metadata:
  • Essay:
    • Device free human activity recognition via wireless signals offers great potential in various consumer markets like smart home or health care. However, applying these recognition systems in new areas (different from the training environment) has been challenging as test-subjects/environments bias them. This week's paper presents EI, a deep learning framework that explicitly tackles this issue by training an activity classifier aimed at filtering out these environment-dependent features. It achieves this by using a convolutional feature extractor with two goals: 1) Aid a classifier that recognizes the current activity correctly and 2) choose features so that a second classifier (discriminator) is not able to identify the environment accurately. This approach enables the model to be trained on mostly environment/subject-agnostic features, thereby making the application in previously unknown environments realistic. The authors demonstrated substantial improvements in activity recognition across four different wireless technologies, especially when looking at generalizability (new settings).
    • This paper once again highlights the power of deep learning models as feature extractors. In some of the shown domains, it would be challenging for humans to find suitable features, yet the algorithm was able to design them internally. It even discarded these that would give too much information about the environment without any human intervention. It is impressive how much potential these approaches hold across various problem domains.
    • Activity recognition based on wireless signals could also be invaluable to privacy-preserving deep learning applications (frequencies are not as personalized as pictures). Enabling those technologies will become increasingly important as awareness for privacy issues continues to rise.
    • Lastly, I think the paper presents the design of EI in a very approachable manner. Even though my exposition to deep learning is limited to introductory courses, I could follow the explanation of their network architecture easily and further improve my understanding of adversarial methodologies.
    • There are two discussions that I would like to see. While they evaluated their approach on four separate types of signals, a combination of multiple might have provided even better accuracy. I think that the correlation between different measured types of frequencies could have served as a feature on its own (either indicating a specific environment or activity). Also, as the paper is another step into a broader adaptation of wireless activity recognition, a comparison with image-based approaches would have been insightful. While they mention that privacy is an issue, I would have welcomed a discussion about the tradeoffs of the methods (e.g. are there issues that wireless is unlikely to overcome?). One could always cross-reference with literature, but having an evaluation of both approaches on the same test data would offer a more solid comparison.
    • On a final note: This paper combined with PA2 made me reconsider my stance on areas like frequency analysis. While I was previously reluctant to dive into electrical engineering related topics as a CS major, I am really starting to see the value of having problem domain knowledge in different areas, especially when working with mobile phone sensors.