Updated the preliminary documentation for Pilot 2 benchmark 2.

Author: bvanessen

P2B2/README.md

@@ -0,0 +1,39 @@
+## P2B2: Autoencoder Compressed Representation for Molecular Dynamics Simulation Data
+
+**Overview**: Cut down on manual inspection time for molecular simulation data
+
+**Relationship to core problem**: Identify possible forming of cancer-causing proteins by molecular simulation
+
+**Expected outcome**: Improvement in the understanding of protein formation and easing of the handling large-scale molecular dynamics output
+
+### Benchmark Specs Requirements
+
+#### Description of the Data
+* Data source: MD Simulation output as PDB files (coarse-grained bead simulation)
+* Input dimensions: ~1.26e6 per time step (6000 lipids x 30 beads per lipid x (position + velocity + type))
+* Output dimensions: 1xN_Frame (N=100 hidden units)
+* Latent representation dimension:
+* Sample size: O(10^6) for simulation requiring O(10^8) time steps
+* Notes on data balance and other issues: unlabeled data with rare events
+
+#### Expected Outcomes
+* 'Telescope' into data: Find regions of interest based on higher level of structure than rest of regions
+* Output range: Dimension: 1 scalar value corresponding to each frame of simulation representing the structured-ness of the data compared to the mean. Output range: [0, 100] 0=mean noise level, 100=very structured.
+
+#### Evaluation Metrics
+* Accuracy or loss function: Domain experts agreeing on utility
+* Expected performance of a naive method: Comparison of different technical approaches against each other and against labels (see above)
+
+#### Description of the Network
+* Proposed network architecture: stacked fully-connected autoencoder feeding RNN
+* Number of layers: 4
+
+### Running the baseline implementation
+
+Using virtualenv
+
+```
+cd P2B2
+workon keras
+python __main__.py  --home-dir=${HOME}/.virtualenvs/keras/lib/python2.7/site-packages --look-back 15 --train --epochs 20 --learning-rate 0.01 --cool --seed --batch-size 10 --seed
+```