Description
Winter 2025
Problem Description: Image Classification with Convolutional Neural Networks
Your goal in this assignment is similar to the goal of the previous assignment (Homework 4) in which you
trained FCNs to classify images in the FashionMNIST data set. In this assignment, however, you will have a
budget of weights that you can incorporate into your neural network model and you will compare FCNs vs.
Convolutional Neural Networks (CNNs).
Figure 1 First 64 Images in FashionMNIST Dataset.
Some comments and hints
Here are some useful comments and facts to guide you along the way.
1. Many of the components of the previous assignment and the setup in the template HW4_HelperTemplate.ipynb
will hold for this assignment as well. Changes that you will need to make will be in the network setup
and input data into your network.
2. The hyperparameter tuning that you performed in the previous assignment should also be useful to
guide you in creating the FCN 100K baseline.
3. You do not need to use a precise number of weights in your network variants. An approximate order of
magnitude will be fine as well (e.g. 87,800 could be considered as 100K).
Tasks
Since this is the last assignment in the course, the definition of the tasks is more open-ended to allow you to
design the neural networks in a less supervised way.
1. With the constraint of incorporating up to 100K weights in the design of your FCN model, perform
hyperparameter tuning to achieve FCN model whose testing classification accuracy is above 88% on
the testing set.
2. Reduce by half the number of weights in the FCN 100K model to create FCN 50K, and also double
the number of weights to create FCN 200K. Train these models similarly to FCN 100K and study the
accuracy of these models by comparing the different FCN variants.
3. Implement and train a CNN 100K model with convolutional, pooling, and FC layers with up to 100K
weights. Perform hyperparameter tuning and compare the testing classification accuracy of the model
with FCN variants.
4. Reduce the number of weights in the CNN 100K model to create CNN 50K, CNN 20K, and CNN 10K.
Train these models similarly to CNN 100K and compare their testing classification accuracy with all
CNN and FCN variants. In addition to accuracy, you may want to include efficiency, in terms of the
number of weights and training time, in your discussion.
5. Bonus (+2 points): Pick a CNN variant from above and then pick several input samples from different
classes. Visualize some of the feature maps of the convolutional layers for these samples (e.g. display
the feature maps in a grid of nxn). Explain your observations.
