A dense network is trained to classify configurations of a polymer chain as coil or…
PHYSICS · Multi-part · Fermi University Physics · Computational Physics · High-performance and data-driven computational physics · High-performance and data-driven computational physics
Scientific Computing for Physicists, Ch Neural networks on physical data: backpropagation, convolution and symmetry
Problem
A dense network is trained to classify configurations of a polymer chain as coil or globule from bead coordinates, with training, validation and test sets separated by independent Monte Carlo run. The table below records the training and validation loss after each epoch. Training loss after epochs 1 to 8 is 0.62,0.45,0.33,0.22,0.14,0.09,0.05,0.02, and validation loss after epochs 1 to 8 is 0.60,0.42,0.31,0.24,0.21,0.23,0.27,0.33. The parameters kept by the rule of part
((a)) are then frozen and evaluated once on the test set, where they score an accuracy of 0.90. (a) State the epoch at which early stopping should keep the parameters, giving the criterion used.
((b)) State the test accuracy that should be reported for this network, and say why the test set should not also be evaluated after every epoch to confirm the choice.