Neural network babysteps with tensorflow.
- Ciresan: http://papers.nips.cc/paper/4741-deep-neural-networks-segment-neuronal-membranes-in-electron-microscopy-images
- U-Net: http://link.springer.com/chapter/10.1007/978-3-319-24574-4_28 / http://link.springer.com/chapter/10.1007/978-3-319-46723-8_49
- FusionNet: https://arxiv.org/abs/1612.05360
- ZNN: https://arxiv.org/abs/1508.04843
- Normalization: How and where? What about BN-Layers?
- Randomization of training data: Is permuting after every cycle the right thing to do?
- U-Net: Where to apply dropout?
- Prunining for NNs start with fully connected NN and drop connections that become too weak. Basically that is L0 (infeasible) / L1 loss <- NN with L1 papers? But could also drop connections once their absolute value becomes too small. Probably with droping rate increasing during the learning process?
- End-to-end learning for connectomics
- DeepQuoran