AI & Machine Learning (KAN-CINTO4003U) - Copenhagen Business School | Fall 2025

---

Group members

Student name	Student ID
#NAME#	#ID#
#NAME#	#ID#
#NAME#	#ID#

---

Mandatory assignment 1

This repository contains the first mandatory assignment (MA1) for AIML25. It consists of two parts: Artificial Neural Networks and Convolutional Neural Networks. Please read the description for each part of the assignment carefully.

• Dev setup: To complete this assignment, you will need to have followed the Development Setup guide. Please complete this setup before starting the assignment.

• Helper code: You are given a lot of boilerplate code as functions in /src to be imported into each of the notebooks. This is to simplify the assignments. Please acquaint yourself with this "helper code".

• Guides: Each of the sub-assignments have a corresponding guide notebook. Everything needed to complete the assignments is in the notebooks. The guide notebooks are there to help you understand and practice the concepts. Please refer to the table below:

Assignment	Assignment notebook	Guide notebook	Description
Part 1	assignments/nn.ipynb	materials/nns_pytorch.ipynb	Artificial Neural Networks
Part 2	assignments/cnn.ipynb	materials/cnns_pytorch.ipynb	Convolutional Neural Networks

Please see due dates and other relevant information in the assignment description on Canvas.

Getting started

Please see the relevant sections in the Development setup guide document on Canvas for instructions. To reiterate, you need to:

1. Fork this repository to your own GitHub account.
2. Clone the forked repository to your local machine.
3. Create a virtual environment and install the required packages.
4. Start working on the assignment.
5. Commit and push your changes to your GitHub repository.
6. Submit a link to your GH assignment repo on Canvas.

Part 1: Artificial Neural Networks

In this assignment, your task is to construct an MLP architecture for the well-known MNIST dataset.

• Read through the material/nns_pytorch.ipynb notebook to understand the basics of PyTorch and how to build a simple neural network.

• Build a n-layer MLP (artificial neural network) in the assignments/nn.ipynb notebook

  • Create the architecture (section 4)
  • Set the training parameters (section 5)

• See how good a model you can build!

Feel free to play around with the model architecture and see how the training time/performance changes, but to begin, you can try the following:

• Input Layer: Image (784 dimensions) ->
• Fully connected layer (500 hidden units) ->
• Nonlinearity (ReLU) ->
• Fully connected (10 hidden units - digits 0-9) ->

Some hints:

• To get the best performance, you may want to play with the learning rate and the number of training epochs.
• A flattened image of size 28x28 is 784 dimensions. This is the input to the network.

---

Part 2: Convolutional Neural Networks

In this assignment, your task is to build a CNN architecture for multi-class classification on the famous CIFAR-10 dataset.

• Read through the material/cnns_pytorch.ipynb notebook to understand the basics of how to build a convolutional neural network.

• Build a n-layer CNN in the assignments/cnn.ipynb notebook.

  • Create the architecture (section 4)
  • Set the training parameters (section 5)

• See how good a model you can build!

Feel free to play around with the model architecture - add, layers, adjust layer parameters etc. - and see how the training time/performance changes. Try to get the best performance you can!

You can start with the following:

• 2D Convolutional layer
• Non-linearity (ReLU)
• Max pooling layer
• Flatten the output from the last convolutional layer (this part is already given in the code)
• Fully Connected Layer with 128 hidden units
• Non-linearity (ReLU)
• Fully Connected Layer (output layer for 10 classes)

Some hints:

• You need to define the number of filters (also known as channels) in each convolutional layer. Try starting with a small number (e.g., 16 or 32) and increase it in deeper layers.
• The kernel size determines how much of the image each filter sees at a time. Smaller kernels (like 3x3) are commonly used and work well.
• Pick a small number for stride and padding. Padding helps preserve the spatial size of the input, and stride determines how much the kernel moves.
• Pooling layers reduce the spatial dimensions of the feature maps and help reduce computation and overfitting. A common choice is to use 2x2 max pooling with a stride of 2, which halves the spatial dimensions.