Building a Deep Learning Framework
Swapneel Mehta (~SwapneelM) |
Ever wondered if you could build your own deep learning framework for hundreds of users? Well, we did build one and turns out it's not as hard as it sounds.
With thousands of people working towards democratising artificial intelligence (AI), we have seen an explosion in the availability of machine learning libraries that make it simpler to build and deploy models for a wide range of tasks. From finance to art, every field has been revolutionised by the introduction of AI.
At the European Organisation for Nuclear Research (CERN) we work on understanding the fundamental particles that constitute the universe by performing various experiments in particle physics. Of late, we have experienced a stratospheric rise in deep learning applications to various problems - RNNs, CNNs, and GANs - that have yielded promising results.
Like, this stuff is so cool. It works!
We delve into the development of one such project as it evolves from a set of scripts into a full-blown framework for supervised learning in high-energy physics.
- In this talk we will detail the evolution on the DeepJet Framework. It will delieate the development isssues, and how it evolved from a set of scripts hastily patched together to a structured, cross-platform framework built on top of Tensorflow and Keras.
The library is a WIP so we're shipping updates on a daily basis with the goal of improving usability with focus on documenting our existing code base.
Initially envisaged to support the development of the namesake jet-tagger in the CMS Experiment at CERN, it has grown to encompass multiple purposes within the collaboration. It is aimed at outlining how to go from a set of scripts to building a library that is used by hundreds of scientists in the world's largest physics research collaboration.
The presentation will describe the major features the environment sports: simple out-of-memory training with a multi-threaded approach to maximally exploit the hardware acceleration, simple and streamlined I/O to help bookkeeping of the developments, and finally Docker image distribution, to simplify the deployment of the whole ecosystem on multiple datacenters.
The talk will also cover future development aimed at improving user experience.
- Experience working with virtual environments or Anaconda
- Knowledge of basic ideas within machine learning such as training, testing, and evaluation of models
Basic knowledge of particle physics helpful but not required
Swapneel is a computer scientist working at Compact Muon Solenoid (CMS) Experiment at the European Organisation for Nuclear Research where physicists and engineers are probing the fundamental structure of the universe. They use the world's largest and most complex scientific instruments to study the basic constituents of matter – the fundamental particles.
His work at CERN encompasses the creation of a framework that can facilitate the use of deep neural networks and provide a suite of functions to serve multiple use-cases such as jet classification, particle identification, and so on. He is an open-source enthusiast, writing and contributing to various projects in his free time.