OnurcanKoken / MasterAI

Master in AI

In this repository, I present some of my works related with Artificial Intelligence at University of Lübeck, during my Master in Robotics and Autonomous Systems. I would like to thank all of my teachers and lab teammates.

Courses: Machine Learning, Computer Vision, Artificial Intelligence, Medical Deep Learning

Machine Learning

1. PCA, ICA and Sparse Coding

   Three unsupervised methods are applied: Prinicipal Component Analysis (PCA), Independent Component Analysis (ICA), and Sparse Coding (SC).

   Dataset: MNIST - handwritten images and faces.

2. Manifold Learning

   Multi-dimensional Scaling and Isomap algorithm are implemented. Dataset: Swiss roll.

3. Statistical Learning Theory

   K-nearest-neighbor algorithm is implemented for a two-class classification problem on a two-dimensional dataset.

4. Boosting

   Adaptive boosting algorithm is implemented to combine multiple weak classification models and form one strong classifier for a two-class classification problem on a two-dimensional dataset.

5. Random Forests

   Random forest algorithm is implemented to combine multiple decision trees to one strong classifier for a multi-class classification problem on a two-dimensional dataset.

Computer Vision

0. Introduction to Python

   • Basic Python: Basic data types (Containers, Lists, Dictionaries, Sets, Tuples), Functions, Classes
   • Numpy: Arrays, Array indexing, Datatypes, Array math, Broadcasting
   • Matplotlib: Plotting, Subplots, Images
   • IPython: Creating notebooks, Typical workflows

1. Imaging - Histograms

   • Image gradients
   • Compute and display histograms
   • Underexposed, darker, images that have been exposed to too little light
   • Overexposed, brighter, images that have been exposed to too much light
   • Histogram equalization, logarithmic and quadratic functions

2. Image Acquisition, Optics

   • Thin lenses, focal length, focus
   • The influence of the focal length
   • The influence of the aperture
   • Conversion between pixel and camera coordinates, Time-of-flight (ToF) camera

3. Image Center, Edges, Keypoints

   • Finding the image center
   • Edge detection, show image gradients
   • Canny edge detector
   • Key point detection by using the Structure Tensor J and the Hassian Matrix H

4. Greedy Snake

   • Use the gradient magnitude as image term
   • Calculate the three energy terms $E_{cont}$, $E_{curv}$, $E_{imag}$
   • Normalize the energy terms over the neighborhood to the intervall $[0, 1]$
   • Estimate the point in the neighborhood, that has minimum energy

Artificial Intelligence - MATLAB

1. Nearest Neigbor

   • Matlab Introduction
   • Brute Force

2. Linear Programming

3. Perceptron

   • Pattern-by-Pattern Learning
   • Batch Learning

4. Quadratic Programming

   • Optimization and Maximum Margin

5. Duality (no code)

   • Dual Linear Program
   • Equilibrium Theorem
   • Strong Duality Theorem
   • Weak Duality Theorem

6. Lagrange Multipliers

7. Polynomial Kernels

8. Simplified Sequential Minimal Optimization (SMO)

9. Support Vector Regression (SVR)

10. Neural Networks

    • Classification
    • Sigmoid Activation Function
    • Multi-Layer-Perceptron (MLP)

Medical Deep Learning - PyTorch

0. Introduction to PyTorch

   • Tensor basic
   • Tensor playground
   • Autograd
   • CNN
   • ConvBlock
   • Pneumonia classification in x-ray torso
   • Dataset
   • Pytorch data tensor structure
   • Training

1. 3D Semantic Segmentation

   Implementation of a deep 3D-CNN for multi-organ segmentation.

   Dataset: Learn2Reg Abdominal CT (20 training and 10 validation scans each manually labelled with 13 organs).

   • Affine transformation
   • 20 training and 10 validation scans each manually labelled with 13 organs
   • Training the FCN for segmentation
   • Segmentation using lite reduced atrous spatial pyramid pooling (LR-ASPP) (missing part)

2. Electrocardiogram (ECG) Sequence Classification

   Classify (variable-length) ECG recordings of the PhysioNet/CinC Challenge into four rythm classes: normal sinus rhythms (N), atrial fibrillation (AF), other rhythms (O), noise signals (∼).

   • PhysioNet/CinC Challenge

   • Preprocessing

     To use the ecg signals with our neural nets normalize each recording s to have zero mean and unit variance.

   • Pre-trained 7 Layer ResNet (featnet) + Linear Classifier

   • 15 Layer ResNet + Linear Classifier

   • 15 Layer ResNet + LSTM + Linear Classifier

   • Transformer

3. Image Registration

   Introduce deep-learning based image registration.

   • Computation of (joint) histograms and mutual information
   • Image Transformation
   • (Discrete) Correlation layer
   • Define feature network architecture
   • Training of MI-based registration
   • Modality Invariant Neighbourhood Descriptor (MIND) (missing part)
   • Learning global + local deformable multimodal registration (missing part)

4. Weakly-Supervised Visualization

   Implement methods that allow to gain insights which parts of an input image to a Deep Neural Network are pivotal for its classification decision.

   • Implement dataset and visualize the given pancreas train & test data
   • Finetune a pretrained ResNet on the given data
   • Implement the Class Activation Mapping (CAM) method
   • Implement the guided backpropagation
   • Implementing the custom ReLU-Layer

5. Model Distillation & Ternary Nets

   Implement methods that allow to compress deep learning models via model distillation and ternary weights. This enables the use of deep learning in medicine due to its real-time ability and implementation on weaker mobile devices.

   Data of Patch Camelyon (tupac16) Challenge is used.

   • Modify a pretrained VGG11_BN network for the given training data
   • Fine tuning
   • Network Pruning through increased Sparsity
   • Ternary weight approximation (missing part)