Autoencoders for sample size estimation for fully connected neural network classifiers

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Sample size estimation is critical in experimental design but remains understudied for deep learning.
• Current methods rely on heuristics or prior experience, often leading to inefficient data labeling for supervised learning tasks.

Purpose of the Study:

• To address the underestimation of sample size requirements in deep learning, particularly for computer vision.
• To develop a rigorous method for estimating the minimum labeled data needed for effective model training.

Main Methods:

• Investigated the concept of a Minimum Converging Sample (MCS) representing the smallest dataset for a generalizable representation.
• Utilized autoencoder loss to estimate MCS for fully connected neural networks in computer vision tasks.
• Developed a code-free, dataset-agnostic tool for MCS estimation.

Main Results:

• Found that below the estimated MCS, fully connected networks struggle to differentiate classes.
• Demonstrated a strong correlation between generalizability and autoencoder loss for sample sizes above the MCS.
• Successfully provided a practical tool for estimating sample sizes.

Conclusions:

• Minimum Converging Sample (MCS) estimation using autoencoder loss is a promising approach for guiding data collection and labeling in deep learning.
• This method can significantly improve the efficiency and effectiveness of training computer vision models.
• The findings offer a more data-driven strategy for sample size determination in deep learning applications.