Question

Computational Intelligence Methods

ii) Describe the process that has been followed to implement a Multi-layer Perceptron for handwritten digit recognition as presented in the lecture (the code is included in the accompanying Jupyter Notebook ipynb file). Experiment by changing the network's architecture (number of neurons in hidden layer and number of hidden layers) and comment on how the accuracy of the network is affected by these changes.

Answer 1

Solution 1: One of the most famous solutions to perform handwritten recognition is through using deep learning architecture. In this, a significantly deep neural network is employed to perform the classification of the digits on the basis of the features present in the digits images. It is more of a computer vision problem that makes use of a very deep learning neural network that is able to learn the trends and the patterns present within the images that are fed to it through a dataset explicitly, on the basis of a suitable learning algorithm such as backpropagation algorithm. It can also be a convolutional neural network that does not require you to actually pass the features of the images explicitly instead it learns the features from the images on the basis of the different regions and the shapes present within the image.

Solution 2: A neural network is a very powerful machine learning model and it can provide exceptional classification as well as the regression capabilities to the projects that someone is actually working on. So a multilayer perceptron is a very strong machine learning architecture and this is because since it has multiple neurons present within the neural network and hence it has the capabilities to learn a lot of parameters from the data on itself and hence it has exceptional generalization capabilities. This makes them a lot more powerful classifiers and regressors as compared to other machine learning algorithm or architecture. So if you have a smaller number of perceptrons then your model may suffer from the underfitting and would have less generalizational capabiltiy while if it has a lot of perceptrons then it can perform the generalization as well as the classification way better than the previous case.

Here's the solution to your question, please provide it a 100% rating. Thanks for asking and happy learning!!