Year: 2024
Type: other
Abstract: Background: Dataset complexity significantly impacts the performance of machine learning (ML) models and is influenced by factors such as dataset size, feature dimensionality, data distribution, and the presence of noises. Evaluation of dataset complexity is crucial to optimize the model’s performance and to anticipate challenges during the process. Various measures, including entropy, fractal dimension, Kolmogorov complexity, and neural network-based approaches like autoencoders, can quantify dataset complexity. Among these, Autoencoders (AE) are found to be a promising method for evaluating dataset complexity, particularly for image datasets. However, the effectiveness of AEs compared to the other measures of dataset complexity is still an open question. Methods: This study investigates the effectiveness of convolutional autoencoders (CAE) in evaluating the complexity of image datasets. Experiments were conducted on different image datasets with varying levels of complexity, using various AE architectures. A methodology based on CAE was developed, using reconstruction loss as a measure of dataset complexity. The relationship between dataset complexity and CAE performance was analyzed to assess the effectiveness of CAE in measuring dataset complexity. Results: The findings indicate that CAEs can effectively evaluate the complexity of image datasets, offering a simpler and less mathematically intensive approach compared to other measures. The study also suggests that increasing the latent space dimension improves the quality of reconstructed images, consistently reduced reconstruction loss for both datasets, and enhances CAE performance. Notably, the validation loss for the ABOships dataset was consistently lower than for the COCO dataset, indicating its lower complexity. Conclusion: Evaluating the complexity of image datasets is critical for developing efficient models for computer vision tasks. The study's findings suggest that CAE can assist in evaluation dataset complexity providing a reliable and efficient approach. However, further research is needed to compare the effectiveness of AEs with other measures of dataset complexity and to investigate their application in various domains. Keywords: Dataset complexity, Convolutional Autoencoder, Reconstructing loss, Latent Space dimensionality
Source: Doria (University of Helsinki)
Author: Vatandoust, Monire
Institution: Åbo Akademi University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7115334027