pseyfert/transfer-learning

an attempt to use arXiv:1409.7495 in data from https://www.kaggle.com/c/flavours-of-physics

transfer learning

an attempt to use arXiv:1409.7495 in data from https://www.kaggle.com/c/flavours-of-physics

transfer-learning / domain adaptation / gradient reversal layer

The authors of arXiv:1409.7495 show their idea of how to train a machine learning classifier on some one domain of data (studio photos / simulated events in HEP) and apply it to another domain of data (smartphone photos / real data in HEP). The framework is Caffe with their fork on github.

flavours of physics

for obvious reasons I have access to simulated Ds→φ(µµ)π events, simulated background events to Ds→φ(µµ)π (i.e. the simulated events used in the training of the classifier for τ→µµµ events in arXiv:1409.8548, but instead of applying the τ→µµµ selection, I use the Ds→φ(µµ)π selection). Furthermore, I have a mix of real events with the Ds→φ(µµ)π selection (with some real events, some background events, and I hardly know which are which).

pack it together

I want to do the following

• train a TMVA classifier (here I know what I'm doing) purely on MC (signal and background) and select in data some events and make a nice invariant mass plot

• train Caffe with only MC events (to see how much performance I gain / lose when using a toolkit I don't know), make an invariant mass plot with the same number of events (see how much cleaner the S/B gets).

• train Caffe with GRL (i.e. labelled MC events and unlabelled data events). Make an invariant mass plot with the same number of events (see how much cleaner the S/B gets). In comparison to the other Caffe network, I will see how much the transferlearning gains in performance.

license

The project code is licensed under the MIT license.

The project logo is from wikipedia and licensed under CC0 1.0.