Assignment 3

For  Assignment 3 you can either complete option 1 or option 2. These options correspond to:

1. An assignment based on implementing batch normalisation for fully connected networks.
2. An assignment based on training a ConvNet to classify the country of origin of a surname.

Advice about the options:

Option 1: I expect most students to choose this option.

Option 2: This assignment is long (at least its description is!) and more fiddly to implement then the other assignments. I suggest that you only attempt this if you are very keen to implement a ConvNet from scratch. Note the simple basic approach proposed in the assignment does not work particularly well given the dataset. However, it does make predictions much better than random and the main point of the assignment is get familiar with a ConvNet implementation.

Option 1: Material for the assignment 

•  Assignment3_option1.pdf  - Instructions for the assignment exploring batch normalisation for fully connected networks.

•  ComputeGradsNumSlow.m - Matlab function to compute the gradients numerically (centred difference method). Note in this function I assume the parameters of the network are stored in the structure NetParams and this includes the mean and standard deviation values if the network has batch normalization layers.

Option 2: Material for the assignment 

• Assignment3_option2.pdf  - instructions for the assignment exploring training a ConvNet to classify a short sequence of letters.
• ascii_names.txt
• category_labels.txt
• Validation_Inds.txt
• ExtractNames.m
• DebugInfo.mat
• NumericalGradient.m
• ascii_names_with_multi_label.txt  - this is only needed for a bonus point assignment