The tutorial:

Deep learning is one of the fastest growing branches in machine learning, due to its spectacular performance in human cognitive tasks.

The neural network has many layers (‘deep’), and its main implementation is through ‘convolutional neural networks (CNNs)’. In 2012 the challenge to classify the images of the ImageNet database with 14 million images was won with a strikingly better performance that earlier methods.

The deep structure of many convolutional layers is also recognized in our human visual perception.

CNNs find applications in scene recognition, self-driving cars, medical diagnosis, translation etc.
The technology is feasible, as today we have abundant computing power, and access to big data.
It is embraced by the biggest companies (Apple, Google, Facebook, Baidu), and is rapidly transforming many areas of our technological society.

A convolutional neural network

This course will give a step-by-step introduction to deep neural networks. We will discuss the terminology of many concepts, study the famous papers by the inventors, and implement our first steps in CNNs on some instructive toy databases, such as MNIST for handwritten digit recognition. As this field is also known as ‘brain-inspired computing’, attention is also paid to models of human visual perception. In the course, many real-world examples will be discussed and explained. We will discuss a number of well-established mathematical modeling techniques in detail, in particular multi-scale and multi-orientation differential geometry, models for self-organization and plasticity, and geometric neural feedback, leading to effective adaptive operations. We present the theory in an axiomatic, intuitive and fundamentally understood way.

The lectures will be in the morning, in the afternoon we practice all concepts in a computer lab.
The course is concluded with a written exam.

This is a short intensive course of three full days, where each morning of lectures is followed by a computer lab in the afternoon, to bring the concepts to life (all software code is supplied).
We learn to design everything interactively with the powerful design language Mathematica version 11 (www.wolfram.com).
Students will work in small groups on the assignments.
 

Time: 30th Aug. 2017 (Wed.) -1nd Sep. 2017 (Fri.) 09:10-12:10 &13:20-17:00 (第2,3,4節&第6,7,8,9節) (Every day)

Location lectures: NTUST TR-212

Location computer lab: NTUST TR-212

Each day consists of 3 lectures of 45 minutes in the morning, and in the afternoon we practice the concepts in a computer lab.

The tutor:

Prof. Bart M. ter Haar Romeny, PhD
Eindhoven University of Technology
Department of Biomedical Engineering
Biomedical Image Analysis
Den Dolech 2 – GEM-Z 2.106
NL-5600 MB Eindhoven, the Netherlands
Tel. +31-40-2475537, mob.: +31-6-24235693
Email: B.M.terHaarRomeny@tue.nl
Homepage: http://bmia.bmt.tue.nl/people/BRomeny/index.html
 

Objectives of the course:

The introduction of deep learning and its applications.
To give a base of some mathematically well funded computer vision methods.
The course will focus on modern and robust applications, in particular on medical imaging.
We will also introduce the notion of geometric reasoning', enabling us to express tasks on images.
This requires some mathematical insight, which will be discussed and explained extensively.
This course invites to ‘play with the math’ during the interactive computerlab sessions.
The methods are inspired by the stunning performance of human visual perception.
We shortly discuss modern findings in brain and visual system research, and how they can be exploited in our algorithms.

The reason we use Mathematica 11 in the computer lab and all lectures, is that is is is a high-level software language
with unequalled possibilities for design, with extensive visualization of the results, and interactive manipulation of parameters.
It integrates symbolics, fast numerics (now faster than Matlab) and excellent graphics.

Lectures overview:

First day: Wednesday 30 August 2017, room NTUST TR-212


Material:

Lecture of this morning: NTUST 2017 Deep Learning Introduction.pptx (PPT)
Deep learning by Geoffrey Hinton (PPT)


12:10-14:00 Lunch


14:00-17:00 Computer lab

Task 1: Study the document An Introduction to Convolutional Neural Networks - Teach.pdf

Task 2: Study the PPT: NTUST 2017 Deep Learning Introduction.pptx (130MB), or pdf (3.4MB)

Task 3: Get to know a little bit of Mathematica 11.

  Convolution01 (notebook)
  Convolution02
  Convolution03

Tutorial Mathematica notebooks:

Course part 1 of 3 (English)
 

Course part 2 of 3 (English)
 

Course part 3 of 3 (English)


       << Benchmarking`;
       BenchmarkReport[ ]

(PS: don't forget the backquote `)

  •  You can inspect if your system is ready for parallel processing (on multiples cores, or on the GPU with CUDA or OpenCL) with:
    SystemInformation[ ]

 

 

Second day: Thursday 31 August 2017, room NTUST TR-212

Rehearsal and discussion of the most important units in a CNN

  •  Convolutional layer

  •  Max and mean pooling layer

  •  Rectifying linear units

  •  Fully connected layer

  •  Error backpropagation and gradient descent learning

  •  Learn to recognize handwritten digits with the MNIST dataset

  •  Data augmentation

  •  Recurrent neural networks

  •  Network visualizations

The human visual system

  •  Learning receptive fields by Principal Component Analysis (PCA)

  •  Physiology of vision

12:10-14:00 Lunch

14:00-17:00 Computer lab

Part I:

Study the PPT by Geoffrey Hinton: Hinton-IntroDeepLearning.pptx

Study in Wikipedia (only the beginning of the pages):

  •  Principal Component Analysis
  •  Eigenvalues and eigenvectors
  •  Covariance matrix

Study the PPT: Physiology of Vision.pptx

Study the PPT: Learning Receptive Fields.pptx

Study the instructive examples of Deep Networks (and the visualizations!) here: http://cs.stanford.edu/people/karpathy/convnetjs/

Part II:

Now we know the basis notions, you can study the profossional (and famous) packages
(this is beyond a course in just 3 days!)

There are many entries into Deep Learning.
The website that gives a really good overview is:

   www.deeplearning.net.

PS: for the exam requirements: see below.


Third day: Friday 1 September 2017, room NTUST TR-212

·        Differential structure of images, shape detection

Convolution and Fourier Transform

Example of a clinical application:
Screening for retinal damage from diabetes by deep learning

12:10-14:00 Lunch

15:00-16:00 Exam

 


Exam

The exam will be on Friday 1 September 2017, 15:00-16:00 in NTUST room T2-511.

The topics you should study:

  •  All concepts from the slides presented.
    •  Examples:
      •  Backpropagation
      •  Receptive field
      •  ReLU
      •  Learning rate
      •  Fully connected layer
      •  Synapse
      •  etc. etc.
         
  •  The following mathematical concepts:
    •  Convolution
    •  Correlation
    •  Covariance
    •  Gradient operator
    •  Principal component analysis
    •  Eigenvector
    •  Invariance
       
  •  Some general notions of Deep Learning
    •  Be able to give application examples of Deep Learning
    •  Know about the limitations
    •  Know examples of Deep Learning languages
    •  Know some examples of classifiers
       

Success!!

 
 

Mathematica 11:

We design all our algorithms in Mathematica 11 (Wolfram Inc., Champaign, Ill.).
Mathematica is a powerful software environment for symbolic and fast numerical computing.

Everything can be made interactive with a single line of code, enabling easy 'playing with complex math'.

The executable of Mathematica 11 can be downloaded from here, or you can download a free trial version from Wolfram's website.
Many universities and institutes offer it as part of their campus license software.

Bring your own laptop and do all experiments directly yourself!


Prof. Bart M. ter Haar Romeny, PhD
Eindhoven University of Technology, Department of Biomedical Engineering
Biomedical Image Analysis
B.M.terHaarRomeny@tue.nl