Construction of a neural network for handwritten digits recognition based on TensorFlow library applying an error backpropagation algorithm
Tetyana O. Filimonova, H.T. Samoylenko, Anna V. Selivanova, Yu.Yu. Yurchenko, Alexei Parashchak · Eastern-European Journal of Enterprise Technologies · 2023
The object of this study is a neural network for recognizing handwritten digits based on the TensorFlow library using the backpropagation algorithm. The main problem addressed is the development of an effective model with high recognition accuracy. Working on such a task is important as it allows understanding how algorithms and models can effectively work with real data and helps improve machine learning techniques. It has been determined that after 20 training epochs, the loss function is 0.105, and the recognition accuracy is 0.976, comparable to human recognition capability. The classification report indicates that the model is effectively trained on training data and demonstrates high accuracy on test data, capable of generalizing information to new examples. Visualization of recognition results confirms that the model correctly recognizes even poorly written digits. The results can be explained by the peculiarities of the model architecture, optimal selection of hyperparameters, and successful use of the backpropagation algorithm, which was not explicitly specified during model training. TensorFlow provided a convenient toolkit for implementing the neural network and optimizing its parameters. As a result, the model has a fairly high accuracy in image recognition. A significant feature of the results is the high recognition accuracy achieved through the optimal model architecture, correct choice of hyperparameters, and effective use of the backpropagation algorithm. Unlike models built using Keras and convolutional layers, the research model quickly learns, which is important, and does not compromise on accuracy. This result was made possible by the above features of model construction. The results could be practically applied in the field of handwritten character recognition, especially in automated document classification systems, in banking recognition systems, and in other areas where the accuracy of handwritten character recognition is essential