Search Papers | Poster Sessions | All Posters
Poster B101 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
The landscape of functional neuron embeddings depends on regularization
Polina Turishcheva1, Viktor Dobrev1, Max Burg1,2,3, Michaela Vystrčilová1, Alexander Ecker1,4; 1University of Goettingen, 2International Max Planck Research School for Intelligent Systems, Tübingen, Germany, 3Tübingen AI Center, University of Tübingen, Germany, 4Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany
Understanding the functional landscape of neurons is crucial for developing a taxonomy of neuronal cell types. Recent work proposed an approach to identify functional cell types by learning a predictive model that approximates the input-output function of a population of neurons and represents each neuron's function by an embedding. These neurons' embeddings have been used to investigate the landscape of cortical computation in the early visual system, but it remains unclear how the structure of the embedding space depends on the design choices of the predictive model. There were two major differences in architectures: (1) a change of spatial sampling strategy for neurons receptive field; and (2) using dynamic video stimuli instead of static images. Here we investigate the impact of such design choices on the functional landscape in the mouse primary visual cortex. We find that strong L1 regularization of the final linear layer, essential for earlier models, is vital for structured embeddings, even with more recent architectures that do not require regularization to achieve strong predictive performance. Varying the backbone architecture did not significantly impact the embeddings structure. Overall, our work is an important step towards interpretable brain modeling and taxonomy of cell types in the visual system.
Keywords: neural response modeling visual cortex representational learning cell types