•  
  •  
 

Abstract

Loss functions are used for estimating the accuracy of the classification models in Machine Learning. The error loss value measures whether the predicted labels match the true labels or are close to them. Handwritten character recognition is a pattern recognition problem. The progression of pattern recognition applications for offline image classification is rapidly improved by using deep learning techniques. Convolution Neural Networks (CNN) and the activation function are the most commonly used in deep learning. In addition, different loss functions are used to evaluate the performance of the model such as Categorical Cross-Entropy (CCE) and Mean Square Error (MSE). In deep learning techniques, the size of the datasets plays an important role in obtaining high performance, but with traditional MSE, the loss values reach zero in the very early stages of the training process when the dataset size is large, and yet the model accuracy is still in progress. This study proposes a developed loss function via enhancing MSE. The proposed improved MSE is based on dividing the square error by the sum of the predicted label probabilities instead of the total of the sample number. Five datasets are used to test the performance of the proposed modified MSE with the proposed CNN pipeline model in addition to the modified VGG16. The datasets are AHCD, AIA9K, HIJJA, Self-collected, and MNIST. The loss rates of the proposed loss function showed a significant improvement in the accuracy rates for all datasets with the improved MSE in comparison to the CCE.

Keywords

Loss functions, Machine learning, Pattern recognition, Handwritten character recognition, Deep learning, Mean square error

Subject Area

Computer Science

Article Type

Article

First Page

1016

Last Page

1033

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Download

Share

COinS