Back
Tailored ensembles of neural networks optimize sensitivity to stimulus statistics
The capability of a living organism to process stimuli with nontrivial intensity distributions cannot be explained by the proficiency of a single neural network. Moreover, it is not sufficient to maximize the dynamic range of the neural response; it is also necessary to tune the response to the intervals of stimulus intensities that should be reliably discriminated. We derive a class of neural networks where these intervals can be tuned to the desired interval. This allows us to tailor ensembles of networks optimized for arbitrary stimulus intensity distributions. We discuss potential applications in machine learning.
@misc{item_3292883, title = {{Tailored ensembles of neural networks optimize sensitivity to stimulus statistics}}, booktitle = {{DPG-Fr\"uhjahrstagung 2021 BP-CPP-DY-SOE}}, abstract = {The capability of a living organism to process stimuli with nontrivial intensity distributions cannot be explained by the proficiency of a single neural network. Moreover, it is not sufficient to maximize the dynamic range of the neural response; it is also necessary to tune the response to the intervals of stimulus intensities that should be reliably discriminated. We derive a class of neural networks where these intervals can be tuned to the desired interval. This allows us to tailor ensembles of networks optimized for arbitrary stimulus intensity distributions. We discuss potential applications in machine learning.}, pages = {47}, year = {2021}, slug = {item_3292883}, author = {Zierenberg, J and Wilting, J and Priesemann, V and Levina, A} }