machine-learning-applications

notes for Machine Learning -- Applications course

timeline and grading

lecture notes:

• machine learning fundamentals

laboratory classes

1. Handwritten digits classification using MNIST dataset with Pytorch

• models: perceptron, deep fully-connected network, generic CNN
• various activations,
• overfitting,
• regularization, early stopping

Colab notebook

2. ECG signal classification

• classifiers comparison: SVM, decision trees, random forests
• feature vectors

Colab notebook

3. Image classification using deep CNNs

• VGG, ResNet

Colab notebook

4. Regularization

• L2 and L1 regularization implemented by hand

Colab notebook

5. Augmentation in image processing, two separated tasks:

• take MNIST or CIFAR dataset, apply some simple geometric transformations (see e.g. lecture), and check if such dataset extending improves accuracy (take some CNN model from previous labs):
  • use simple transformations (e.g. flip, rotate, translate, scale) using scikit-image, or open-cv
  • or use albumentations library, demo: https://albumentations-demo.herokuapp.com/

  • example of combining albumentations with pytorch Dataset is presented here

  • in case of MNIST verify if applying flips or rotations > 45 deg improve accuracy or not, why?
• play with one-shot style transfer that might be also used for images augmentation (e.g. see here), understand the idea, and run some exemplary code on your own images

• papers:
  • Gatys original paper
  • https://arxiv.org/pdf/1904.11617v1.pdf
• code:
  • https://github.com/limingcv/Photorealistic-Style-Transfer
  • https://github.com/KushajveerSingh/Photorealistic-Style-Transfer

6. Convolutional GAN on MNIST

• generative adversarial network model: generator & discriminator
• training GANs

Colab notebook

7. TBA

proposed seminars topics

• list of proposed topics
• link to form with seminars dates