Ryan Peters

One of the fundamental goals of neuroscience is to model the relationship between neural activity and behavior. With recent advances in technologies for coupling large-scale neuronal recordings alongside high-fidelity behavioral recordings, there is an increased demand for more advanced models to study these population-level dynamics. This research project demonstrates the use of CEBRA, a novel contrastive learning based approach for non-linear dimensionality reduction of neuronal populations recently developed at the Mathis Laboratory of Adaptive Intelligence, as applied to 1-photon calcium recordings of a trained GCaMP-6f mouse navigating through a figure-8 maze environment. We demonstrate that, across three sessions, CEBRA produces stable embeddings that are representative of the physical figure-8 environment. We also demonstrate the stability of these population-level embeddings across multiple sessions within the context of behavioral decoding, and compare CEBRA’s performance against the traditional Bayesian decoder. Finally, we provide evidence that latent embeddings exhibit enhanced structural coherence when trained with neurons from the retrosplenial and visual cortex in comparison to other brain regions, such as the motor cortex.