Can neural networks count digit frequency?

TL;DR

This paper compares classical machine learning models and neural networks in counting digit frequency within numbers, finding neural networks outperform classical models in accuracy and generalization across multiple datasets.

Contribution

It introduces a hybrid classification-regression approach for digit frequency counting and systematically evaluates different models on custom datasets.

Findings

Neural networks outperform classical models in accuracy.

Decision trees and random forests overfit and do not generalize well.

Neural networks show significant improvements in both regression and classification tasks.

Abstract

In this research, we aim to compare the performance of different classical machine learning models and neural networks in identifying the frequency of occurrence of each digit in a given number. It has various applications in machine learning and computer vision, e.g. for obtaining the frequency of a target object in a visual scene. We considered this problem as a hybrid of classification and regression tasks. We carefully create our own datasets to observe systematic differences between different methods. We evaluate each of the methods using different metrics across multiple datasets.The metrics of performance used were the root mean squared error and mean absolute error for regression evaluation, and accuracy for classification performance evaluation. We observe that decision trees and random forests overfit to the dataset, due to their inherent bias, and are not able to generalize well. We also observe that the neural networks significantly outperform the classical machine learning models in terms of both the regression and classification metrics for both the 6-digit and 10-digit number datasets. Dataset and code are available on github.