Function and organization of an autoassociative olfactory memory network

Olfaction is an important sensory modality for many animals, including humans. The neural computation converting the chemical information of odors into a meaningful representation in the brain is under active research. While the representations reflect the chemical composition of odors, they are also significantly shaped by the animal's individual experience. The first part of this thesis explores how experience influences the odor representations, using zebrafish olfactory system as a model.
We focused on the brain region known as Dp, a homolog to the mammalian piriform cortex, which is proposed to function as an autoassociative memory network.
Calcium signals were recorded from the pDp subregion of Dp in juvenile zebrafish, which were either naïve or trained in an odor discrimination task. Activity in the olfactory bulb (OB) from the same fish were subsequently recorded, providing information of input signals to the pDp. Combining this dataset with odor-evoked activities recorded in adult zebrafish, we found that traditional network models predicting point attractor dynamics fail to explain several observations. Instead, a model that liberated activity patterns from being confined to a single point attractor, and allowed them to spread into a subspace with enriched structure, better described the data and revealed the effect of experience on odor representation through geometrical measures. The subspace with enriched structure is abstracted as a neural manifold representing an odor. We implemented and computed distance metrics between manifolds, and applied the framework of manifold capacity to quantify the separability of the manifolds. These geometric analysis captured the transformations of manifolds induced by learning. The results argues that rather than creating discrete memory items, pDp and related recurrent networks make use of complex and continuous geometric transformations of item representations to store behavior-relevant information.
In the olfactory system, as in many other sensory systems, computation is distributed across multiple brain regions among a large number of neurons. Especially, the OB and Dp interacts to process the olfactory signal through the bottom-up and top-down pathways. The function and structural organization of these bi-directional pathways are not fully understood. The second part of this thesis describes an effort to map the connectivity between the OB and Dp at single-neuron level in a juvenile zebrafish. A high-resolution 3D image stack of an OB and the ipsilateral telencephalon was acquired with the serial block-face scanning electron microscopy (SBEM). The image tiles in the stack were carefully aligned and ground-truth segmentation were generated to prepare for future semi-automated neuron segmentation. Moreover, the stack was taken from a brain sample of a fish trained in an odor discrimination task and the odor-evoked calcium activities in the OB and Dp were recorded. Correlating the somata in the EM stack to those in the calcium images identified the neurons with both structural and functional characterization. Although the final steps of analysis on this stack remains for future work, the project is part of the establishment of the functional connectomics pipeline in the lab.
The two projects studied the autoassociative olfactory memory network in Dp from the functional and structural perspectives, and together provide insights into the mechanisms of olfactory computation and olfactory memory formation.