Automatic Differentiation and Deep Learning

• Go to Indico Event
• Where: 222/R-001 at CERN

Statistical learning has been getting more and more interest from the particle-physics community in recent times, with neural networks and gradient-based optimization being a focus.

In this talk we shall discuss three things:

• automatic differention tools: tools to quickly build DAGs of computation that are fully differentiable. We shall focus on one such tool "PyTorch".
• Easy deployment of trained neural networks into large systems with many constraints: for example, deploying a model at the reconstruction phase where the neural network has to be integrated into CERN's bulk data-processing C++-only environment
• Some recent models in deep learning for segmentation and generation that might be useful for particle physics problems.