Deep learning for imaging calorimeters - datasense talk

Transcript

1. Deep learning for imaging calorimeters Mehdi Cherti Supervisor : Balázs Kégl (Linear Accelerator Laboratory / AppStat group) 1 DigiCosme/Datasense, July 8th 2014
2. Outline • Introduction • Calorimeters • Description of the project • Convolutional neural networks • Results and discussion • Perspectives 2
3. Introduction (1/2) 3 Context : International Linear Accelerator (ILC) project Purpose : Measure empirically the efficiency of deep learning algorithms on raw calorimeter data
4. Introduction (2/2) 4 • Mainly to avoid Feature engineering Why deep learning?
5. Calorimeters (1/2) • A calorimeter is an energy measurement tool used in particle physics experiments 5
6. Calorimeters (2/2) 6
7. Description of the project (1/4) • 4D data : For each position (x, y, z) we have an energy deposit • Input dimensionality : 9720 pixels 7 18 18 18 18 30 30 x y z
8. Description of the project (2/4) • Baseline : we use manual extracted features and we train them with adaboost (multiboost software) • These features are : front image, layers difference and layers energy 8
9. Description of the project (3/4) 9 x Front image : y z x y covariance matrix of the points (x, y)
10. Description of the project (4/4) 10 x z y Layers difference Layers energy z :
11. Convolutional neural networks • Supervised deep learning algorithm • State of the art results in object recognition • Detect features regardless of their position with convolution • Some translation invariance with pooling 11
12. Results and discussion (1/3) • Pre-processing (standardization) and Rectified linear units (ReLU) made the optimization step faster • Pooling did not help much 12
13. Results and discussion (2/3) 13 Epoch Epoch
14. Results and discussion (3/3) • Convolutional networks did not help much but we should point out that the dataset is noisy • Convolutional networks are less prone to overfitting 14 Noisy instances:
15. Perspectives • Separating tracks of two independent particles (mixture of particles) • Prediction of the energy and the number of secondary particles • Prediction of the energy of the incoming particle • Design a generative model for calorimeter data 15 Mixture of particles :
16. Thank you for your attention.
    ! Questions? 16