Sensing the Past with Smell

When an object or person leaves the scene, visual, auditory, and linguistic cues disappear instantly. Smell is the only modality that lingers, allowing us to look into the past. We study this new perception problem by developing a model that can use smell to sense the past. Building on recent work in collecting large-scale odor datasets and training machine-learning classifiers for real-time olfactory perception, we collect SmellPast, capturing odors after their sources are removed for up to 2 hours. We show that substances can be classified with up to 50% accuracy (on 5-way classification over 4 substances and no substance) 2 hours after removal. We introduce a Decay Feature GRU that augments sensor readings with their relative patterns across sensors and their rates of change, improving recognition over a standard GRU at every evaluated time from 10 minutes to 2 hours after removal. These results suggest a novel way of sensing the physical world, extending current work from sensing across space to sensing across time.