Visualizing What our Layers are Learning
- In order to implement neural style transfer, we need to evaluate the features at various layers in our convolutional network
- In order to do this, we can visualize what our layers (and neurons) are actually learning
- To do this, we should: 1. Train a network 2. Pick a unit in each layer 3. Find the nine image patches that maximize the unit's activations
- In other words, we should find the region of an image that maximize the activations returned by a particular neuron
- This will help us in understand what each neuron is actually learning
- As a reminder, the deeper layers will learn more specific details about an image
- On the other hand, the shallow layers will learn more generalized information about the image
- This is because neural networks learn moving from the back of the network to the front of the network (i.e. backward propagation)
- In other words, a convolutional network breaks the image down moving from higher level to lower level layers
Describing Visualize
- define feature map
- visualizing network using max activations
- visualizing network using mean activations
tldr
References
Previous
Next