[x] is the improvemet with batch normalization really a surprise? With BN you need to fit one parameter less (as z is determined by x and y now). So the network needs to learn less stuff. And probably also good to have a limited output range. Otherwise the solution would also not be unique as 1,1,1 and 2,2,2 would yield the exact same prediction score.
[x] you often showed the slide step by step. In general it is better for these kind of talks to show the full slide right away.
[ ] specify how exactly you calculated the SNR. Probably best to directly change the x-axis label to a more descriptive name.
[x] make Rayleigh fit to see how distribution fits our expectation.
[x] the average resolution is meaningless as we average over a non representative energy distribution.
[ ] check SNR histogram
[x] CHange to \sigma_68 instead of I_68 in report + plots etc.
I have a few takeaways/comments