Home
Course Guidelines
About the course Prerequite Material References
Python
Jupyter Notebooks Python overview
Exercises
Before the semester start: Installation and exercise setup Week 1: Introduction to Python and libraries Week 2: Vector representations Week 3: Linear Algebra Week 4: Linear Transformations Week 5: Models and least squares Week 6: Assignment 1 - Gaze Estimation Week 7: Model selection and descriptive statistics Week 8: Filtering Week 9: Classification Week 10: Evaluation Week 11: Dimensionality reduction Week 12: Clustering and refresh on gradients Week 13: Neural Networks Week 14: Convolutional Neural Networks (CNN's)
Tutorials
Week 1: Data analysis, manipulation and plotting Week 2: Linear algebra Week 3: Transformations tutorial Week 4: Projection and Least Squares tutorial Week 7: Cross-validation and descriptive statistics tutorial Week 8: Filtering tutorial Week 11: Gradient Descent / Ascent
In-class Exercises
In-class 1 In-class 2 In-class 10 In-class 3 In-class 4 In-class 8
Explorer

Document

  • Overview
  • 1. Optimization
  • 2. NN architectures
  • 3. Bias variance and regularization

Content

Convolutional Neural Networks (CNN's)

In the following exercises, Convolutional Neural Networks (CNNs) will be used for multiclass classification with the PyTorch library. The objective is to explore CNN architectures, focusing on how network topology and optimization strategies influence performance. The performance of the CNN will be compared to a Multilayer Perceptron (MLP) to determine which architecture is better suited for the Fashion MNIST dataset. Additionally, the concepts of regularization and cross-validation will also be examined.

Learning goals

After this week's exercises, you should be able to:

  • Experiment with the CNN architecture to better reason about the impact of the different factors involved.
  • Visually investigate different layers and reason about the effect of different layers on the prediction.
  • Evaluate the performance and reason about optimization strategies.
  • Evaluate the performance and reason about tuning techniques.
  • Evaluate the performance and discuss advantages and limitations.