Unsupervised Domain Adaptation in Smart Buildings.

What is it about?

In this work, we consider unsupervised adversarial domain adaptation (UADA) to estimate the occupants’ number and recognize their activities. Adversarial learning develops a discriminator to distinguish between samples from source and target domains and a feature extractor to fool this discriminator which creates better domain-invariant feature representations or it learns transferable feature representations by embedding an adversarial learning module within a deep learning model. We adapt 4 UADA approaches called Drop to Adapt (DTA), Drop to Adapt with Virtual Adversarial Training (DTA+VAT), Batch Spectral Penalization with Domain Adversarial Neural Network (BSP+DANN), and Batch Spectral Penalization with Conditional Domain Adversarial Network (BSP+CDAN). These adapted approaches have been trained and evaluated on sensor data for AR and OE for both balanced and unbalanced label proportions. The outstanding performances with scores up to 98% prove the efficiency of the adapted unsupervised adversarial domain adaptation methods. The complete source code is in the following repository: https://github.com/JawDri/Unsupervised-Adversarial-Domain-Adaptation-for-OE-and-AR.git.

Featured Image

Photo by Markus Spiske on Unsplash

Photo by Markus Spiske on Unsplash

Why is it important?

In the last two decades, smart buildings have profited from AI-based approaches to create smart applications that provide several advantages such as energy efficiency and security. In particular, occupancy estimation (OE) and activity recognition (AR) are vital solutions that study the frequency and behavior of occupants within a building. The smart building labeled data is scarce, and its collection is costly, tedious, time-consuming, and can be infeasible due to privacy issues. Researchers have considered unsupervised domain adaptation (UDA) techniques to solve the problem of data scarcity by sharing knowledge from source domains where labeled data is abundant to target domains where labeled data is not available. In previous research, domain adaptation approaches learn invariant feature representations from both source and target domains to allow knowledge transfer across domains.

Perspectives