A Study of Artificial Neural Network and Its Implementation from Scratch

Abstract

Nilay Kushawaha, Ankhi Roy

One of the major problems in the field of artificial intelligence (AI) is the use of machine learning model as a black box, even though it might be helpful in a few cases but understanding the internal structure and the operating mechanism will assist the user to tweak the variables in a more efficient and productive manner. In this paper we have introduced the working of an artificial neural network (ANN) by taking the example of a three layered neural network. The entire mathematics behind the working of neural network along with the different evaluation metrics required to assess the performance of the model are discussed in this paper. We have also created a custom neural network from scratch and compared it with the keras based model on three different datasets - susy dataset [8], cardiovascular dataset [6], churn dataset [7]. The results obtained demonstrates that the overall performance of both the models are almost identical which gives an idea that it is possible to train a neural network from scratch without the use of any framework.

PDF