The framework thus offers a path towards circuit models—for olfactory sensing and beyond—that both perform well in naturalistic environments and make experimentally-testable predictions for neural response dynamics.

Overall, our model separately infers odor concentration and presence, achieving faster and more robust inference. At the same time, our model is itself a recurrent circuit that demonstrates rich cell-type-specific neural dynamics resembling those that have been observed in the OB.

Two simulations were developed: one quantifying the model’s inference ability on a timescale of hundreds of milliseconds, while the other examining how the required number of OSNs scales with the size of the potential odorant dictionary.

Lastly, we evaluate how our model scales with increasing network size and odor dimensionality and how its performance varies with different affinity matrices and priors.

Next, we mapped the model’s inference dynamics on the circuit architecture of the olfactory bulb. Notably, SDEO naturally gives rise to two classes of projection neurons resembling mitral and tufted cells and providing experimentally testable predictions.

We then built a biologically plausible recurrent neural network implementing these sampling dynamics. Through simulations, we demonstrated that our SDEO accurately tracks the presence and concentration of changing odorants.

MLD performs sampling in an unconstrained dual space and projects samples back to the constrained primal space via an invertible mirror map, therefore obtaining samples from a constrained distribution.

To enable rapid inference of binary odor presence in a biologically plausible recurrent network, our model leverages the framework of Mirror Langevin Dynamics (MLD).

We proposed “SDEO”, a model for olfactory compressed sensing inspired by simultaneous localization and mapping (SLAM) algorithms in navigation: the set of odors that are present in a given scene, and the concentration of those present odors, are inferred separately.

Thrilled to share our new preprint, now on bioRxiv!! Huge thanks to all collaborators!

For those interested, here’s a bit more about the work:

The framework thus offers a path towards circuit models—for olfactory sensing and beyond—that both perform well in naturalistic environments and make experimentally-testable predictions for neural response dynamics.

7/7 b

Our model, which separately infers odor concentration and presence, performs faster and more robust inference of odorants. At the same time, our model is itself a recurrent circuit that demonstrates rich cell-type-specific neural dynamics resembling those that have been observed in the OB.

7/7 a

Two simulations were developed: one quantifying the model’s inference ability on a timescale of hundreds of milliseconds, while the other examining how the required number of OSNs scales with the size of the potential odorant dictionary.

6/7 b

Lastly, we evaluate how our model scales with increasing network size and odor dimensionality and how its performance varies with different affinity matrices and priors.

6/7 a

Next, we mapped the model’s inference dynamics on the circuit architecture of the olfactory bulb. Notably, SDEO naturally gives rise to two classes of projection neurons resembling mitral and tufted cells and providing experimentally testable predictions.

5/7

We then built a biologically plausible recurrent neural network implementing these sampling dynamics. Through simulations, we demonstrated that our SDEO accurately tracks the presence and concentration of changing odorants.

4/7

MLD performs sampling in an unconstrained dual space and projects samples back to the constrained primal space via an invertible mirror map, therefore obtaining samples from a constrained distribution.

3/7 b

To enable rapid inference of binary odor presence in a biologically plausible recurrent network, our model leverages the framework of Mirror Langevin Dynamics (MLD).

3/7 a

We proposed “SDEO”, a model for olfactory compressed sensing inspired by simultaneous localization and mapping (SLAM) algorithms in navigation: the set of odors that are present in a given scene, and the concentration of those present odors, are inferred separately.

2/7