Design and develop two feedforward neural network using two different configurations as given below. You can use any number of layers [2, 3, 4] and any number of neurons [128, 256, 64, 32, 16] in each layer. You are given 20 newsgroup datasets for classification. The dataset high-level details are given below. You must submit your source code (python notebook) along with your answer for this question.

 

Classes	20
Samples total	18846
Dimensionality	130107
Features	real

 

1. Compute f1, precision, recall and accuracy to evaluate your feedforward neural classifiers. 
2. Compare performance of your classifiers by plotting a single bar graph showing all four metrics for two classes. Please write down the reasoning behind performing better or worst performance for each classifier. 
3.  Show case (plotting a line curve) and explain the effect of different learning rate effects of your best performing classifier designed above. Assume learning rate varies from [10, 1, 0.1, 0.05, 0.01, 0.0015, 0.001] for a fixed number of epochs. 
4.  Assume you have 70-30 split of your dataset, where 70% data (13193 samples) used for training and 30% data (5653 samples) used for testing. Randomly select 5, 10, 20, 30, 40, 50 % of your training data and change the actual labels with incorrect labels (you can do it randomly) and train your best performed network for this different noisy training dataset. Evaluate your noisy models and compare performance using the metrics (acc, f1, precision, recall). Explain your reasoning in detail why performing better or worst. If any trend exists in the performance due to adding noise in the training data, then explain the reason in detail?