• A simple image classification implementation, where description and classification are done together.
• Models used
  • Linear single-layer perceptron.
  • Non-linear single-layer perceptron.
  • Multi-layer perceptron.
  • A pipeline with a convolutional layer.
• All models, forward and backward propogation steps are implemented from scratch.
• This is unlike a shallow classification, where description and classification are done separately.

Problems in hw4.pdf are solved.

• All source code is in cnn.py & main_functions.py. Ironically cnn.py is the main file.
• main_functions.py reads mnist_train.mat & mnist_test.mat.
• Description of data format is given in hw4.pdf.
• comps contains results of models for different hyper parameters. It mainly serves as an example of how NOT to do a hyper parameter search :)

• Code is the documentation of itself.

• Use python3 cnn.py to classify images and visualize results using above mentioned models.
• A summary of the methods and corresponding results is given in report.pdf.

• Linear single-layer perceptron.

• Non-linear single-layer perceptron.

• Multi-layer perceptron.

• A pipeline with a convolutional layer.