Classification of heart disease using MLP with SeLU activation function and Binary Cross Entropy as Loss function

Abstract

Multilayer perceptron are the basic components of deep learning and neural net- work characterized by their architecture of feedforwading consisting of fully con- nected neurons with the non linear activation functions.This thesis provides a comprehensive overview of MLPs,by giving detailed about their structure and com- ponents,including the input layer,hidden layers, and output layer. This explains the role of weights and bias neurons in the process of learning and adjustment of the complex pattern in data. Different activation functions which are Sigmoid, ReLU, tanh, were used, illustrating their important in non-linearity and using the model to find its relationship.The training of MLPs by using the back propagation method is thoroughly studied. This includes the forward pass,where the input data is stored and processed though the network,and the backward pass, where gradients of loss function are calculated to update the network parameters. The importance of the loss estimation and parameter updates are used in minimiz- ing the error and improving the model’s performance.Additionally, we will discuss about the important steps in data preparation, which involves the handling missing values,feature scaling, and ensuring the proper data formatting. The importance of splitting the data into training and testing is highlighted, Technique for preventing the over fitting, such as regularization,dropout, and early stopping. The guidelines for using the MLP is provided, starting with the simple one and then increasing the complexity based on different tasks.This thesis focuses on the usage of exper- imentation with various architecture,hyper parameters and optimizing the model to get the optimal results. Different techniques used for monitoring the training and the adjustment of hyper parameters are discussed, along with the methods to calculate the model accuracy and loss checking and test datasets.This thesis aims to adopt practitioners with the understanding of tools to develop robust and efficiency of model on different applications such as image recognition,NLP, and speech recognition. By following the detailed guidelines and using the best tech- niques in data preparation, training and evaluation can increase the performance of model.