Weekly meeting 13 Dec 2016

Meeting minutes for weekly phd meeting 13 Dec 2016:

==================================

I showed Tom the curves of inference speed versus image input size.

Ideas for the future:

============

CNN architecture:

1) Scaling (s)

2) Residual blocks of convolutions

3) Upsample (1/s)

4) Loss layer with two parts: dice coefficient, inverse of the scaling factor.

Compression:

sensitivity based iterative pruning of the residual convolutions.