We hope our focus on fluctuating odors will inspire more experiments with naturalistic temporal statistics in olfactory mixtures. On the theory side, the manifold learning concept may inform background filtering strategies in other biological and artificial systems.

We then dissect, numerically and analytically, the dynamics and stability of these models (detailed Appendix!). Notably, we find that local interneurons endowed with the IBCM learning rule decompose the background by selecting individual odors.

We combine biologically plausible models of synaptic plasticity to implement this strategy in the early olfactory circuit. Simulating these learning rules show they greatly improve the recognition of new odors in fluctuating backgrounds.

Our proposal: manifold learning is optimal against fast fluctuations in high-dimensional olfactory space. Instead of filtering by anticipating inputs, a better strategy is to learn the low-dimensional manifold of background odors and suppress input projections in that subspace.

Due to turbulent transport, natural odor signals are intermittent and vary rapidly. Yet, animals filter out backgrounds with such fluctuations to discern new, relevant olfactory cues. How does that work?

Congrats @vmochulska.bsky.social and Paul, so nice to see this story officially published! The figures look amazing!

Merci beaucoup Paul pour la nomination, la supervision et le soutien pendant ma thèse!