GANspire is a deep learning tool that generates expressive breathing waveforms for art and health applications.
Our design intention for GANspire is threefold. First, we wanted to create a deep generative model that captures fine-grained, expressive variations of breathing in humans. Second, we wished to design a tool that facilitates exploration of the generative model for non-expert users. Third, we sought to apply our tool to art and design projects exploring breathing as a creative material, such as the RÉESPIRATION art-health research project.
We led participatory design with respiratory care practitioners to develop such a deep learning tool, that we named GANspire. We crafted a deep learning prototype by training a WaveGAN model over a small dataset of breathing pressure waveforms sampled in physiology studies. We extracted control parameters from the breathing latent space using the GANSpace algorithm. We developed an interface for exploration of breathing generation using the Marcelle web-based toolkit. The resulting tool helped us probe generative deep learning for respiratory care.
The project was developed with Baptiste Caramiaux, Thomas Similowski, and Samuel Bianchini in collaboration with the HCI Sorbonne group of ISIR, the R3S department of AP-HP, and the Reflective Interaction group of EnsadLab, in the context of a postdoctoral fellowship program at Faculty of Medicine, Sorbonne Université (IUIS).
Paper at NeurIPS 2021