Omni-modal decomposition autoencoders learn full-stack wearable disentangled representations
The episode reviews a paper on OmniDecVAE, a model that handles wearable sensor data from up to 30 channels with a single shared encoder and decoder. It achieves strong activity and identity recognition, generation, and disentanglement in one framework, with constant parameter count and low latency, though it underperforms in subject-independent settings.