# Artiphysiology project
## Project proposal
### Aim

In this project, we try to better understand an anatomically constrained convolutional neural network (MouseNet) by different visulaization techniques. Specificaly, we would like to explore

- Compared to pretrained artificial networks in the literature, is there any special property in the MouseNet?
- How do properties in the visualization emerge during training? How much training accuracy is enough to get that property?
- How large are learned receptive field size?

### Visualization techniques
- Gradient visualization
- Filter visualization
- Receptive field mapping

### Stimulus
Cifar10 and ImageNet dataset
### Target network
Anotomically constrained MouseNet with the following structure trained on Cifar10 and/or ImageNet dataset.
In this figure, every arrow represent a combination of Conv/BN/Relu/Sparsity_mask with kernel size labeled on the arrow. The first and second number in yellow circle represent the output size and number of output channels in that layer.