sandyfloren / final-nn

Implement a generic neural network class and use it for autoencoding and classifying.

Final project: neural network

Overview

In this assignment, you will implement a neural network class from (almost) scratch. You will then apply your class to create both:

(1) a simple 64x16x64 autoencoder.

(2) a classifier for transcription factor binding sites.

You will begin by finishing the API for generating fully connected neural networks from scratch. You will then make Jupyter Notebooks where you create, train, and test your autoencoder and classifier.

Step 1: finish the neural network API

For steps 2 and 3

• Finish all methods with a pass statement in the NeuralNetwork class in the nn.py file.

For step 3

• Finish the sample_seqs function in the preprocess.py file.
• Finish the one_hot_encode_seqs function in the preprocess.py file.

Step 2: make your autoencoder

Background

An autoencoder is a neural network that takes an input, encodes it into a lower-dimensional latent space through "encoding" layers, and then attempts to reconstruct the original input using "decoding" layers. Autoencoders are often used for dimensionality reduction.

Your task

You will train a 64x16x64 autoencoder on the digits dataset. All of the following work should be done in a Jupyter Notebook.

To-do

• Load the digits dataset through sklearn using sklearn.datasets.load_digits().
• Split the data into training and validation sets.
• Generate an instance of your NeuralNetwork class with a 64x16x64 autoencoder architecture.
• Train your autoencoder on the training data.
• Plot your training and validation loss by epoch.
• Quantify your average reconstruction error over the validation set.
• Explain why you chose the hyperparameter values you did.

Step 3: make your classifier

Background

Transcription factors are proteins that bind DNA at promoters to drive gene expression. Most preferentially bind to specific sequences while ignoring others. Traditional methods to determine these sequences (called motifs) have assumed that binding sites in the genome are all independent. However, in some cases people have identified motifs where positional interdependencies exist.

Your task

You will implement a multi-layer fully connected neural network using your NeuralNetwork class to predict whether a short DNA sequence is a binding site for the yeast transcription factor Rap1. The training data is incredibly imbalanced, with way fewer positive sequences than negative sequences, so you will implement a sampling scheme to ensure that class imbalance does not affect training. As in step 2, all of the following work should be done in a Jupyter Notebook.

To-do

• Use the read_text_file function from io.py to read in the 137 positive Rap1 motif examples.
• Use the read_fasta_file function from io.py to read in all the negative examples. Note that these sequences are much longer than the positive sequences, so you will need to process them to the same length.
• Balance your classes using your sample_seq function and explain why you chose the sampling scheme you did.
• One-hot encode the data using your one_hot_encode_seqs function.
• Split the data into training and validation sets.
• Generate an instance of your NeuralNetwork class with an appropriate architecture.
• Train your neural network on the training data.
• Plot your training and validation loss by epoch.
• Report the accuracy of your classifier on your validation dataset.
• Explain your choice of loss function and hyperparameters.

Grading (50 points)

Neural network implementation (15 points)

• Proper implementation of methods in NeuralNetwork class (13 points)
• Proper implementation of sample_seqs function (1 point)
• Proper implementation of one_hot_encode_seqs function (1 point)

Autoencoder (10 points)

• Read in data and generate training and validation sets (2 points)
• Successfully train your autoencoder (4 points)
• Plots of training and validation loss (2 points)
• Quantification of reconstruction error (1 point)
• Explanation of hyperparameters (1 point)

Classifier (15 points)

• Correctly read in all data (2 points)
• Explanation of your sampling scheme (2 points)
• Proper generation of a training set and a validation set (2 point)
• Successfully train your classifier (4 points)
• Plots of training and validation loss (2 points)
• Reporting validation accuracy of the classifier (1 point)
• Explanation of loss function and hyperparameters (2 points)

Testing (7 points)

Proper unit tests for:

• _single_forward method (1 point)
• forward method (1 point)
• _single_backprop method (1 point)
• predict method (1 point)
• binary_cross_entropy method (0.5 points)
• binary_cross_entropy_backprop method (0.5 points)
• mean_squared_error method (0.5 points)
• mean_squared_error_backprop method (0.5 points)
• sample_seqs function (0.5 points)
• one_hot_encode_seqs function (0.5 points)

Packaging (3 points)

• Installable module (1 point)
• GitHub Actions (installing + testing) (2 points)

Submission

Please submit a link to your final project repo here