Research Computing Summer School 2019

 

The Research Computing Service, the Computational Methods Hub, the Research Software Community and the Graduate School are happy to present the Research Computing Summer School 2019.  This event will bring together the scientific community, local experts and external lecturers for three days of tutorials, lectures and exchange of ideas.

There has been a large demand for training in machine learning across the College and this year's summer school is dedicated to the topic.  We will have a two-day tutorial on the fundamentals of machine learning, one half-day tutorial on medical image analysis, keynote and research talks on applications of machine learning from local experts.

 

Dates: September 25-27, 2019
Venue: Huxley 342 and 308  (building 13 on the map)
Campus: Imperial College London, South Kensington campus
Organizers: Graduate SchoolComputational Methods HubResearch Software Community and the Research Computing Service
Contacts: Katerina Michalickova and Jeremy Cohen
Registration: Please register here for the Machine learning application showcase.

 

 

Programme:

Wednesday, September 25
Time  Huxley 308
10:00-17:00

Fundamentals of machine learning, Marc Cohen and Polong Lin, Google [info]

THIS SESSION IS NOW FULLY BOOKED.

Thursday, September 26
Time Huxley 308
10:00-17:00

 

Fundamentals of machine learning, Marc Cohen and Polong Lin, Google [info]

THIS SESSION IS NOW FULLY BOOKED.

Friday, September 27
Time Huxley 308
10:00 - 13:00

Medical image analysis tutorial, Ben Glocker and Nick Pawlowski, Department of Computing [info]

THIS SESSION IS NOW FULLY BOOKED.

14:00 - 15:50

Register for the afternoon session.

Machine learning applications showcase -  5 x 20 minute presentations:

Deep Reinforcement Learning with PyTorch [abstract]   
Kai Arulkumaran
, Dept of Bioengineering

Generative adversarial networks for visual synthesis and beyond
Dr Viktoriia Sharmanska, Dept of Computing

How machine learning will re-define the state-of-the-art in climate and weather forecasting [abstract]
Dr Peer Nowack, The Grantham Institute for Climate Change

Applications of AI to security and energy efficiency and nature [abstract]
Dr Stephen McGough, School of Computing Science, Newcastle University

Earth and medical imaging: learning from machine learning [abstract]
Professor Mike Warner, Dept of Earth Science & Engineering

16:00 - 17:00

Keynote:

From the Big Bang to AI

Professor Roberto Trotta, Department of Physics, Director of Centre for Languages, Culture and Communication

Cosmology has made spectacular progress in the last 20 years: thanks to telescopes and space-borne observatories, we have been able to map out the history of the cosmos from a fraction of a second after the Big Bang to today, 13.8 billion years later. However, this has led to some of the most fundamental questions in the whole physics, which remain to this day unanswered: What are dark matter and dark energy, which together make up for 95% of the universe? What drove the exponential expansion of the universe it its early stages? Are there other universes out there? As the amount of available data grows exponentially, our ability to answer these questions in a data-driven way rests on novel, faster and more ingenious data analysis and statistical techniques. Machine Learning and AI are quickly becoming the most powerful tools to investigate the fundamental nature of the cosmos.

17:00 - 18:30

Reception

 

 

 

Fundamentals of machine learning

Workshop prerequisites:  Knowledge of Python (Python Tour or equivalent) is desirable.
Set-up instructions:  Please bring a fully charged laptop with a browser.
Slides including many useful links.

DAY 1 (10:00 - 17:00, hourly breaks)

Day 1 morning (10:00 - 13:00)

  • Introduction
    • Welcome
    • Introductions
    • Agenda
  • Foundation
    • What is Deep Learning (DS vs. AI vs. ML vs. DL)? [MC]
    • Industry Examples of AI [PL]
    • Overview of ML terminology [PL]
    • ML Taxonomy
    • Mathematical Building Blocks (What is a tensor?) [MC]
    • Tensorflow vs. Pytorch [PL]
    • What we’re going to build
  • Hello Tensorflow!
    • What is Google Colab? (vs. Jupyter Notebooks vs. other IDEs?)
    • Intro to Tensorflow and Keras
      • Hello World example 

Day 1 afternoon (14:00 - 17:00)

  • In-Depth Examples
    • Binary Classification - movie review sentiment assessment [PL]
    • Multi-class Classification - tagging news articles [PL/MC]
    • Regression Analysis - predicting house prices [MC]

 

DAY 2 (10:00-17:00, hourly breaks)

Day 2 morning (10:00-11:30)

  • Diving Deeper
    • Anatomy of a Model: Recognizing Handwritten Digits [MC]
    • More advanced example: bit.ly/mco-mnist-lab, with more stepwise refinement
    • Concepts
      • Training, Validation, and Testing
      • Preprocessing and Feature Engineering
      • Overfitting and Underfitting
      • Dropout, Batch Norm, and Pooling
      • Scaling computational resources

Day 2 afternoon (14:00 - 17:00)

  • ML in the Cloud [MC]
    • What is the Cloud?
    • What’s happening now at Google?
    • Pre-trained Models
    • AI Platform
    • AutoML
  • ML Workflow
    • How to productionize/deploy models in the real world
    • Tools: Colab vs. Jupyter vs. JupyterLab vs. scripts and apps
  • Wrap-Up, Feedback Survey, and Raffle [MC]

 

 

Medical image analysis tutorial 

Tutorial slides and repo.

Machine learning and deep learning techniques are increasingly being used within the medical imaging community. They can help to undertake tasks that may be challenging or time consuming to carry out manually, and help to extract information from images that may not be obvious to the human eye. They can also offer huge increases in performance for large-scale image analysis tasks that are important in research studies and the processing of large imaging databases. DLTK, Deep Learning Tool Kit (https://dltk.github.io/), is an example of an open source toolkit that builds on TensorFlow to provide implementations of a range of methods that can be used to help automate medical image analysis processes.
 
Using such tools requires some baseline knowledge of the medical imaging domain and core image analysis techniques. This tutorial session will focus on providing you with background knowledge to enable you to further explore more advanced learning-based techniques for medical image analysis. It will also provide an introduction to some imaging-focused machine learning techniques.
 
Prerequisites: 
The tutorial assumes a basic working knowledge of Python and the use of Jupyter notebooks. We will make use of the SimpleITK Python library.
 
Outline:
 
  •  Installation of SimpleITK via pip or conda
  •  Overview of imaging file formats and how to load 3D medical images
  •  Basic visualisation and extraction of image metadata and statistics
  •  Applying filters
  •  Introduction to image segmentation techniques
  •  Image-based predictive modelling

 

 

Deep Reinforcement Learning with PyTorch

Deep reinforcement learning (DRL) is the study of learning to act optimally in sequential decision problems through trial and error (reinforcement learning), using artificial neural networks (deep learning) to scale to problem domains with structured, high-dimensional inputs, such as images and text. In this talk I will briefly introduce reinforcement learning, discuss some theoretical and practical applications of my research in DRL, and finally talk about some of my open source work with PyTorch, which is regularly utilised by both academic and industrial researchers.

 

 

How machine learning will re-define the state-of-the-art in climate and weather forecasting

The potential of machine learning to revolutionize climate science, meteorology and other areas of environmental science is just starting to be realised. In my talk, I run through a diverse range of machine learning applications to central environmental research challenges from my group at Imperial, from speeding up global climate models run on supercomputers to improving weather and air pollution forecasts.

 

 

Applications of AI to security and energy efficiency and nature

Machine learning can be applied in many areas using many different approaches. In this talk I shall present some different application areas and how we have used machine learning to solve specific challenges. In the first I shall show how a single image of a Ransomware screenshot can be used to train a Deep Learning network capable of identifying the variant of ransomware that is displayed on your monitor. In the second I shall demonstrate how Reinforcement Learning and Genetic Algorithms can be used to reduce the energy consumed in a high-throughput computing system. Finally I shall show how Deep Learning can be used to tell the difference between individuals in a set.

 

 

Earth and medical imaging: learning from machine learning

Lluis Guasch, Jiashun Yao, George Stronge, Oscar Bates & Mike Warner, Departments of Earth Science and Engineering, and Bioengineering

Earth scientists use seismic waves to image the interior of the Earth, and medical scientists use ultrasound waves in an analogous way to image the interior of the human body.  Parallel developments in both fields have recently seen computationally intensive wave-equation-based optimisation methods improve resolution and fidelity significantly beyond that which can be attained by time-of-flight tomography and conventional imaging.  The computer codes used in these new methods have much in common with the convolutional networks often employed in deep machine learning.  We are exploring three ways in which this similarity may be exploited:  In the first, we are using theory, algorithms and heuristics developed within the ML community to improve both the efficiency and effectiveness of the optimisation schemes used in earth and medical imaging.  In the second, we are using generative adversarial networks in various configurations to modify the data, intermediate results and models with which conventional optimisation operates with consequent improvements in efficiency and outcomes.  And in the third and most challenging, we are seeking to replace conventional optimisation with a purely ML-based approach, obtaining physical models directly from observed data via ML.  This last approach may not prove feasible for generic Earth imaging since the necessary volume of training data is unlikely ever to be available, but it does seem feasible in some areas of medical imaging where there are both very large datasets available and where all human bodies prove to be much more similar than do all regions of the Earth.  Earth scientists and medical practitioners then can benefit from machine learning, not only by applying ML directly to their problems, but also by adopting into conventional optimisation some of the features that ML specialists have already learnt about the best ways to train and apply their ML networks.