2. Tutorial

We use the AlexNet model and BOLD5000 dataset 1 to demonstrate the versatility and usability of DNNBrain in characterizing the DNN and in examining the correspondences between the DNN and the brain. The results are also used in our paper, and more description about the model and dataset can be found at the ‘Methods’ part of this paper.

Data:

All the data you’ll need for this tutorial is uploaded to OSF. Here are some descriptions and specific download links.
The stimulus images are available by clicking here, and a .stim.csv file named as all_5000scenes.stim.csv is used to tell DNNBrain where and what the inputs are.
The BOLD response maps for each image are available at here.
The VTC mask is available by clicking here.
More details about the data can be found in readme.txt.

Tutorials:

  • In Scan DNN tutorial, we extract and display feature maps of three images for each convolutional layer after ReLU.

  • In Probe DNN tutorial, we reveal animate information presented in DNN layers.

  • In Map between DNN and brain tutorial, we examine how well the representation from each layer predict the response of a voxel in the brain by using voxel-wise encoding models. In addition, we also use representational similarity analysis to characterize the link between the representations of DNN and brain.

  • In Visualize DNN tutorial, We use three visualization ways to examine the stimulus features that an artificial neuron prefers.

References:

1

Chang, N., Pyles, J. A., Marcus, A., Gupta, A., Tarr, M. J., and Aminoff, E. M. (2019). BOLD5000, a public fMRI dataset while viewing 5000 visual images. Sci. data 6, 49. doi:10.1038/s41597-019-0052-3.