Loading…
CNS*2020 Online has ended
Welcome to the Sched instance for CNS*2020 Online! Please read the instruction document on detailed information on CNS*2020.
Back To Schedule
Monday, July 20 • 9:00pm - 10:00pm
P218: The Geometry of Spatio-Temporal Odorant Mixture Encoding in the Drosophila Mushroom Body

Log in to save this to your schedule, view media, leave feedback and see who's attending!

Feedback form is now closed.
Aurel A. Lazar, Tingkai Liu, Chung-Heng Yeh
Google Meet: meet.google.com/rtn-ajej-kuj

Biological organisms are constantly challenged with navigating odorant scenes comprised of complex time-varying mixtures of volatile compounds. To characterize odorant mixture encoding and processing, two seemingly contradictory hypotheses have been considered: Elemental and Configural. The Elemental scheme [1,2] encodes mixtures linearly with identifiable components, while the Configural scheme [3] encodes mixtures as a holistic odor object distinct from its components. Here, we advance a feedback normalization model of the Drosophila early olfactory system that reconciles the two encoding schemes, and analyze the geometry of the resulting odorant encoding space.

Our model consists of Projection Neurons (PNs), Kenyon Cells (KCs) and the Anterior Paired Lateral (APL) neuron. To quantify the degree to which a mixture is encoded elementally vs configurally, we employ the Cosine Similarity (CS) between the KC code of the odorant mixture and its pure components. We show that, due to the global feedback gain control exerted by the APL neuron and the KC spiking mechanism, the steady-state KC output is an input-invariant sparse combinatorial code with consistently 5-10% active neurons. This sparse code results in a configural mixture code with low CS scores against all pure components, enabling the association of different valences to an odorant mixture from its components. Preceding the steady-state phase, the circuit makes full use of gradient encoding in the first two layers of the olfactory pathway [5] and the APL temporal dynamics to encode each mixture elementally with about 25% active neurons. This code exhibits a high (∼ 1) CS score with all pure components in the mixture, indicating high linear decodability. Moreover, we demonstrate that the elemental encoding phase enables cognitive functions for odorant processing. For example, combined with an attention-driven modulation signal, elemental encoding overtakes configural encoding in steady state and promotes odorant tracing for navigation.

Next, we investigate the geometry of the odorant and the KC spaces and show that smooth interpolation between odorant input vectors leads to sharp discontinuities in the KC representation space. Further analysis reveals that the steady-state KC combinatorial codes are almost binary, concentrating around the corners of a high dimensional cube in KC space.

This sparse grid-like structure gives rise to a distinctive clustering of odorant mixture identities in the KC space, with high intra-cluster similarity and inter-cluster dis-similarity. This geometric view of the KC encoding space suggests that sharp transitions in the KC representation is the result of crossing cluster boundaries, which leads to large jumps across vertices of the KC cube, and explains previous observations that small compositional changes of mixtures (mixture ratio or component identities) can incur large differences in perception. Similarly, the transition from elemental to configural phases across time corresponds to a trajectory in the KC space from a subspace spanned by KC codes of odorant components to the vertices of the cube.

Acknowledgments

The research reported here was supported, in part, by NSF grant #1544383 and in part by AFOSR grant #FA9550-16-1-0410.

References

[1] Rokni et al., ACS Chem. Neuro. 2014.

[2] Sehdev et al., Front. Beha. Neuro. 2019.

[3] Thomas-Danguin et al., Front. Psych. 2014.

[4] Dasgupta et al., Science. 2017.

[5] Kim et al., eLife 2015.

Speakers
avatar for Tingkai Liu

Tingkai Liu

PhD Candidate, Electrical Engineering, Columbia University



Monday July 20, 2020 9:00pm - 10:00pm CEST
Slot 08