08 5th February, Monday

I'm currently investigating as to why what I tried on Friday/Saturday failed completely. I think my best bet is to look more towards research that has been done in this area to give me some pointers. I've been reading the original paper on Layer-Wise Relevance propagation, and they discuss the importance of conserving layer-wise relevance when doing the LLR calculation. I've going to continue investigating this to see whether LLR can help be directly, or whether looking at their approach to doing calculations based on weight connections/activations can help me see where I have gone wrong.

Evening update:

I've implemented LLR for my minimal example, using a simplification of the MNIST dataset (only 0s and 7s). Although this has given me some interesting heat maps, I don't feel any closed to solving the problem I had initially set out to solve.

I feel as though this has been fairly unproductive even though I have managed to get a lot of things working. Tomorrow I am planning on changing direction, I feel as though trying to re implement these attacks myself on compressed neural networks would be far more productive than trying to find if I can detect the capacity abuse myself directly based on activations/weights. Hopefully in doing this I can start to produce something that works, and so start moving forwards.