This repository consists of various concept based notebooks in Deep Learning.
To build a deep MLP model for regression task on prediction of house value using California Housing dataset. It introduces the Keras Tuner library that is used to perform hyper-parameter tuning on the neural network. Using the Keras tuner, the best model searched through HyperBand Tuner had 25.13% decrease in the test error compared to the baseline model.
To predict the survival of passengers on Titanic by implementing neural network from scratch using Python incorporating adding multiple layers, different activation functions, feed forward, MSE loss calculation and back propagation, and early stopping to avoid overfitting and comparing its results with voting, average and weighted average ensemble methods using Logistic Regression, Gradient Boosting and MLP classifiers.
The results showed NN with 1 hidden layer showed a Kaggle Public score of 0.79186 which is 2.4% higher compared to NN with 3 hidden layers and ~2% higher compared to ensemble models.