%0 Chart or Table
%A Kiwitz, Kai
%A Schiffer, Christian
%A Spitzer, Hannah
%A Dickscheid, Timo
%A Amunts, Katrin
%T Filter Activations of Convolutional Neuronal Networks Used in Cytoarchitectonic Brain Mapping
%I EBRAINS
%M FZJ-2020-04963
%D 2020
%X We studied the internal structure of two specific Convolutional Neural Networks (CNNs) which were trained to segment the primary (hOc1, V1) and secondary visual cortex (hOc2, V2) in microscopic scans of brain tissue sections with a resolution of 1 micrometer. All tissue sections correspond to those of the BigBrain dataset ([Amunts et al., 2013](https://science.sciencemag.org/content/340/6139/1472)). To analyze the internal feature representations learned by the model, 5184 filter activations from the batch-normalized output of each Rectified Linear Unit (ReLU) of the CNNs were calculated for section number 1021 of the histological stack. We described and analyzed these filter activations to better understand the internal feature representations of the trained networks. This enables a direct comparison with the underlying histology and a direct assessment of cytoarchitectonic features reflected inside the networks. **Additional information:** The corresponding reference delineations were published in: Kiwitz et al. (2019) [DOI: 10.25493/3GSV-T4A](https://kg.ebrains.eu/search/instances/Dataset/87c6dea7-bdf7-4049-9975-6a9925df393f) Kiwitz et al. (2019) [DOI: 10.25493/8MKD-D77](https://kg.ebrains.eu/search/instances/Dataset/02b56db7-a083-44f3-91dc-72bb67f3fd0a)
%F PUB:(DE-HGF)32
%9 Dataset
%R 10.25493/Z6NG-4MU
%U https://juser.fz-juelich.de/record/888499