05 (M) 1st February, Thursday

Meeting today, we discussed the idea I proposed yesterday and a method of generating the vectors that might represent the configuration of the last layer. By taking each of the columns in the matrix that represents the connections between the final layer and the one before, we can generate a number of vectors that could be mapped using t-SNE into the same space as the points from the sampled images. However, this probably won't mean exactly what I'm hoping it will and so it was suggested that I aim to design a small network with a more basic dataset in order to better understand the relationship between the weights of connections in the network and the activation in the final layer caused by different images.

Goals proposed for the week were:

• Try to implement the method discussed to see if anything interesting can be seen.
• Create a minimal network and dataset to test some of the ideas, including trying all possible inputs and mapping these into the lower dimensional space.
• Rerun the last weeks experiment on larger networks and for longer using BlueCrystal, and make sure that the Then/Lasagne framework runs as the tensor flow one does on BC.