## Understanding CNN MouseNet Iris Shi June 2020 <!-- Put the link to this slide here so people can follow --> <!-- slide: https://hackmd.io/@zjjxsjh/Syfcim-IU --> --- ## Introduction of mousenet  <font size="4"> - Each cortical layer (layer 4/23/5) in each area is modelled as a conv layer - Layer 5 from all areas output to a two-layer fully connected classifier </font> --- #### MouseNet is trained on CIFAR10 dataset (32x32, 10 classes)  --- ## Visualization Stimuli  <font size="4"> - Tiny Imagenet test images - 10000 64x64 natural images (downsampled from ImageNet) </font> --- ## Visulization Methods - Gradient visualization with guided backpropagation (GBP) - Direct optimization (Optimization) - both adapted from [pytorch-cnn-visualizations](https://github.com/utkuozbulak/pytorch-cnn-visualizations) by Utku Ozbulak --- ### Conv1 (input->LGNv) all 5 channels  <font size="4"> - Learned features in first layer is simpler than AlexNet - The fist channel is the most complex feature </font> --- ### Conv2 (LGNv->VISp4) all 22 channels  <font size="4"> - The first channel in first conv layer is highly represented by second conv layer with slightly different forms </font> --- ### Conv3 (VISp4->VISp2/3) "best" channels  --- ### Conv3 (VISp4->VISp2/3) "worst" channels  --- ### Visualization during learning  <font size="4"> - Lower frequency features are learned first and then refined to higher frequency features - Lower layer feature emerge earlier than higher layer </font> --- ### Summary and future direction <font size="4"> - Visualization methods only work relatively well on the first three conv layers, possibly due to - GBP: not using enough stimuli - using more stimuli, e.x. various sizes of tiling of ImageNet - Optimization: not enough regularization - adding stronger regularization prior - Intermediate layers are more distributed codes - studying groups of channels instead of single channel </font> --- ### Thank you! :sheep: :100: :muscle: :tada: