Transfer Learning approach for analysis of epochs on Handwritten Digit Classification

In the broader area of machine learning, deep learning depicts a dramatic twist in enhancing the level of intelligence in the various machines. There is a plethora of parameters like weights, bias, number of hidden layers, activation function, and hyperparameters that are utilized for measuring the accuracy of a training model. The prominent form of hyperparameters that are essential for the training process and extracting information for taking decisions is known as Epochs. The enhancement of results is obtained by practicing on the pre-trained network architecture that includes GPU-based computation in the TPU chips. The paper uses the ResNet50 architecture network to train the ultra-deep neural networks for implementing multiclass image classification that is based on Transfer Learning uses a various number of hidden layers to classify handwritten digits and accuracies at different classes is judged by epochs. The tradition of the MNIST database dataset is that it classifies the image into 10 classes. As the image dataset in the NIST is of type bilevel and that was normalized in dimensions to fit in a 224x224 pixel box for better performance. The normalization algorithm uses the anti-aliasing technique to obtain the resultant grey images for the best recognition rate of 99%. As in Traditional machine learning approaches such as KNN, DNN, and many more that take a prolonged time for training the dataset while transfer learning provides the accuracy of 99%, and that was achieved in a few epochs.