Linear NeuroPaint: An Efficient Baseline for Cross-Session Neural Alignment and Inpainting

Linear NeuroPaint: A Comprehensive Evaluation of a Linear Model for Cross-Session Neural Alignment and Inpainting

Large-scale Neuropixels recordings provide insights into various brain areas, yet no single experimental session captures every relevant area. NeuroPaint, a transformer-based autoencoder, addresses this issue by aligning area-specific latent dynamics across sessions and inpainting activity in unrecorded regions. However, could a computationally efficient and potentially more interpretable linear model be sufficient to achieve the same objectives?

To explore this possibility, researchers introduce Linear NeuroPaint, a linear variant of NeuroPaint, and evaluate its performance on two multi-area Neuropixels datasets.

Linear NeuroPaint demonstrates strong predictive performance on recorded areas, indicating its potential effectiveness in this aspect. Nonetheless, when it comes to cross-area inpainting, the linear model falls short compared to its nonlinear counterparts, despite extensive hyperparameter optimization efforts. The reduced accuracy in cross-session alignment and prediction is attributed to incomplete alignment, which results in heterogeneous response profiles among parameter-defined neuron clusters.

Despite these limitations, Linear NeuroPaint serves as a valuable computationally efficient baseline for benchmarking multi-session neural models. Its simplicity and reduced computational demands make it an attractive option for researchers seeking a more interpretable alternative to nonlinear models. However, the lack of effectiveness in cross-session alignment and cross-area prediction highlights the need for further advancements in linear models to fully realize their potential in this domain.