Raphael's presentation - basic neural network (625 -> 256 -> 256 -> 2) for different PTJ values: best results are around 82% categorical accuracy in the 1150-1200 range.
Also, we used a Linear Discriminant Analysis (LDA) algorithm, which is a more general case of the Fisher Linear Discriminant, to check the results reported by Kagan in [1]. At the 250-300 range the accuracy is around 73%.
Jose's presentation - Multilayer Perceptron (2 hidden layers with 5 neurons each): best results are a around 85% also in the 1150-1200 PTJ range. In this model we also verified that it is easier to classify cases in higher PTJ ranges that are visually difficult to classify.
Suggestions for next steps for all.
- Investigate why we get better accuracy on cases that are visually more difficult to classify;
- Use the LDA algorithm for other PTJ ranges;
- Try the same analysis with a convolutional neural network;
Thiago suggestions:
- Find ways to present the results regarding the true positives (plots showing the cases in which the classifier correctly answers signal);
- Better explain the ROC curves used.
- Thiago has to implement the tau21 classifier
