3 papers accepted at NeurIPS ‘23

Our latest work yields state-of-the-art predictions of primary visual cortex (V1) activity in mouse and monkey, plus a new way to optimize stimulation protocols for visual prostheses.

The lab had 3 papers accepted at NeurIPS ‘23:

• PhD students Aiwen Xu and Yuchen Hou developed a multimodal recurrent neural net that well describes V1 activity in freely moving mice, revealing how some neurons lack pronounced visual RFs and that most neurons exhibit mixed selectivity:

  A Xu, Y Hou, CM Niell, M Beyeler (2023). Multimodal deep learning model unveils behavioral dynamics of V1 activity in freely moving mice. 37th Conference on Neural Information Processing Systems (NeurIPS) ‘23

• The latest work by PhD students Galen Pogoncheff and Jacob Granley enriches ResNet50 (the previously best V1-aligned deep net) with layers that simulate the processing hallmarks of the early visual system and assesses how they affect model-brain alignment:

  G Pogoncheff, J Granley, M Beyeler (2023). Explaining V1 properties with a biologically constrained deep learning architecture. 37th Conference on Neural Information Processing Systems (NeurIPS) ‘23

• And last but not least, Jacob Granley (in collab w/ Tristan Fauvel & Matthew Chalk from Sorbonne University) combined deep stimulus encoding with preferential Bayesian optimization to develop personalized stimulation strategies for neural prostheses:

  J Granley, T Fauvel, M Chalk, M Beyeler (2023). Human-in-the-loop optimization for deep stimulus encoding in visual prostheses. 37th Conference on Neural Information Processing Systems (NeurIPS) ‘23