Presentation
(PP28): Evolutionary Convolutional Neural Network for High Energy Physics Detector Simulation
SessionProject Poster Presentation
Event Type
Project Poster
AI/Machine Learning/Deep Learning
Networks
Visualization & Virtual Reality
TimeWednesday, June 19th3:15pm - 4pm
LocationBooth N-230
DescriptionHigh Energy Physics relies on Monte Carlo (MC) for different aspects of data analysis. MC simulation implement complex computations that, today, result in ~50% of CERN Computing Grid resources. Several alternative approaches are being investigated trading some accuracy for speed. Deep Learning approaches, for example, can result in x3000 speed-up while retaining 10% accuracy with respect to MC.
Design and optimisation of neural networks is very difficult and time consuming. Evolutionary computing could be exploited to improve this process, while providing the benefit of both network training and topology optimization in one step. By coding a neural network into genes and chromosome, Genetic Algorithms (GA) would exhibit numerous advantages: easily parallelizable, they could help reach global instead of local optima often selected by Stochastic Gradient Descent (SGD); they could also be combined to SGD for fine tuning.
Our project intends to use GA to train and optimise a 3-dimensional Convolutional Generative Adversarial Network we have developed for particle detector simulation. It is implemented in a few steps: a first phase, will implement a simplified version of our GAN network, focusing on one of the two networks that constitute the GAN (the “discriminator”). We will design a first prototype and define the method applicability (in terms of network complexity and computing resources). We will gradually increase the complexity of the problem, in terms of input image size and network features and finally implement the full adversarial training approach. We will compare performances to standard SGD and classical optimisation approaches (i.e. Bayesian).
Design and optimisation of neural networks is very difficult and time consuming. Evolutionary computing could be exploited to improve this process, while providing the benefit of both network training and topology optimization in one step. By coding a neural network into genes and chromosome, Genetic Algorithms (GA) would exhibit numerous advantages: easily parallelizable, they could help reach global instead of local optima often selected by Stochastic Gradient Descent (SGD); they could also be combined to SGD for fine tuning.
Our project intends to use GA to train and optimise a 3-dimensional Convolutional Generative Adversarial Network we have developed for particle detector simulation. It is implemented in a few steps: a first phase, will implement a simplified version of our GAN network, focusing on one of the two networks that constitute the GAN (the “discriminator”). We will design a first prototype and define the method applicability (in terms of network complexity and computing resources). We will gradually increase the complexity of the problem, in terms of input image size and network features and finally implement the full adversarial training approach. We will compare performances to standard SGD and classical optimisation approaches (i.e. Bayesian).
Poster PDF