1 Next Step in Deep Learning for High

2 Background • Armed Forces – Sig. Int. ; • Undergraduate @ Derby; •

3 Research objectives • Apply Deep Reinforcement Learning to HEP – optimising architecture; •

4 Deep Learning • Artificial neural networks with many hidden layers of neurons; •

5 Machine Learning at LHC • Boosted Decision Trees Scaling leads to over-fitting; Limited

6 Convolutional Neural Networks • State-of-the-art in image classification; • Jet images; • Deep

7 Reinforcement Learning • Learn to behave in an environment through experience • Action

Deep Learning Detector Monte Carlo simulation The entire process is already a form of

9 Monte Carlo simulations • PYTHIA • HERWIG • SHERPA Different MC generators work

10 Training time Depending on the size of the network, training times for deep

Detector MC Generator Pre-processing Improving training times can lead to realtime Deep Learning CNN

12 Training Deep Learning algorithms • High-dimensional error space • Exploiting topological properties •

13 Thank you Questions. . . www. derby. ac. uk/engtech In partnership with

Slides: 13

Download presentation

1 Next Step in Deep Learning for High Energy Physics Craig Bower “Once you stop learning, you start dying” - Einstein www. derby. ac. uk/engtech In partnership with

2 Background • Armed Forces – Sig. Int. ; • Undergraduate @ Derby; • PGCE @ Derby; • Teaching High School; • MSc Comp. Maths @ Derby. www. derby. ac. uk/engtech In partnership with

3 Research objectives • Apply Deep Reinforcement Learning to HEP – optimising architecture; • Reduce training time of Deep Learning algorithms with the aim of real-time implementation at the LHC; • Improve interpretability of Deep Learning models for Physics analysis at ALICE; www. derby. ac. uk/engtech In partnership with

4 Deep Learning • Artificial neural networks with many hidden layers of neurons; • Hierarchical learning of representations; • Low-level input variables; • Physics-engineered variables; • Combination. www. derby. ac. uk/engtech In partnership with

5 Machine Learning at LHC • Boosted Decision Trees Scaling leads to over-fitting; Limited ability to detect complex features; Inputs are high-level physics-engineered variables. • Artificial Neural Networks Example by Mini. Boo. NE experiment, B. Roe et al. , NIM 543 (2005) 577 Limited ability to approximate; Cannot learn features at multiple levels of abstraction. www. derby. ac. uk/engtech In partnership with

6 Convolutional Neural Networks • State-of-the-art in image classification; • Jet images; • Deep correlation images. www. derby. ac. uk/engtech In partnership with

7 Reinforcement Learning • Learn to behave in an environment through experience • Action and reward • Q-function www. derby. ac. uk/engtech In partnership with

Deep Learning Detector Monte Carlo simulation The entire process is already a form of Reinforcement Learning Source: CERN https: //home. cern www. derby. ac. uk/engtech In partnership with 8

9 Monte Carlo simulations • PYTHIA • HERWIG • SHERPA Different MC generators work with different physics models Specific network architectures will be more suitable for different problems www. derby. ac. uk/engtech In partnership with

10 Training time Depending on the size of the network, training times for deep learning algorithms can take days or even weeks. Depends on navigating highly non-convex highdimensional error space. www. derby. ac. uk/engtech In partnership with

Detector MC Generator Pre-processing Improving training times can lead to realtime Deep Learning CNN with Reinforcement Learning Theoretical understanding of model training Deep Correlation Jet Images Physicists Physics Interpretation www. derby. ac. uk/engtech In partnership with 11

12 Training Deep Learning algorithms • High-dimensional error space • Exploiting topological properties • Real-time analysis in preparation for O 2 www. derby. ac. uk/engtech In partnership with

13 Thank you Questions. . . www. derby. ac. uk/engtech In partnership with