CSCI4390-6390 Assign7

Assign7: RNNs

Due Date: Nov 15, before midnight

We will use the Spoken Arabic Digits dataset. This dataset comprises a training and testing dataset. Each dataset records different utterances of digits from 0 to 9 in Arabic. For each digit, 13 features are measured at each time point, though different digits have different number of time points for each sequence. That is, each person has a $\tau \times 13$ sequence for a given digit. Different persons are separated by a blank row in the data files. For a given person, each row is a feature vector of length 13, and the next $\tau$ rows (until the next blank or end of file) represent the sequence for that person.

Also, in the training data the first 660 persons all utter '0', the next 660 utter '1', and so on, until digit '9'. Thus, for each sequence there is only one output label, namely the digit being pronounced.

In the testing data, the format is the same, except each digit has 220 persons, so the first 220 persons utter '0', next 220 utter '1', and so on. See Spoken Arabic Digit Documentation for more details.

Your task is to implement the RNN algorithm 26.1 (on page 679) from scratch in numpy, though you should use the scipy.special.softmax function rather than your own, since it is more robust.

There is one deviation from Algo 26.1. Since there is only one class label per input sequence, there is only one final output from time $\tau$, as discussed at beginning of lecture 20. Therefore, you should adjust the forward and backwards steps in lines 13, 18, and 19, as required.

You should use cross-entropy loss to predict the digit being spoken. The output layer should use softmax, whereas the hidden layers should use ReLU.

Note that sequences are of different lengths, but is easier to just process them one by one (batch size of one). There is no need to try to pad them to be of the same max length, as long as you only do the forward and backward steps only for the actual sequence length $\tau$.

Train on the testing set, and report both the training accuracy and average cross-entropy loss, and then finally report the testing accuracy and loss.

What to submit

  • Submit your python notebook, that contains the solution and output, suitably annotated/commented. Name the notebook: assign7.ipynb.

  • Submit the notebook file via submitty. Your code must be self-contained, and must not hard code file names. You can assume the data files lie in the local dir.

Policy on Academic Honesty

You are free to discuss how to tackle the assignment, but all coding must be your own. Please do not copy or modify code from anyone else, including code on the web. Any students caught violating the academic honesty principle will get an automatic F grade on the course and will be referred to the dean of students for disciplinary action.