Autoencoders for sample size estimation for fully connected neural network classifiers
Abstract Sample size estimation is a crucial step in experimental design but is understudied in the context of deep learning. Currently, estimating the quantity of labeled data needed to train a classifier to a desired performance, is largely based on prior experience with similar models and problem...
Main Authors: | , , , , , , |
---|---|
Format: | Article |
Language: | English |
Published: |
Nature Portfolio
2022-12-01
|
Series: | npj Digital Medicine |
Online Access: | https://doi.org/10.1038/s41746-022-00728-0 |
_version_ | 1797641544927281152 |
---|---|
author | Faris F. Gulamali Ashwin S. Sawant Patricia Kovatch Benjamin Glicksberg Alexander Charney Girish N. Nadkarni Eric Oermann |
author_facet | Faris F. Gulamali Ashwin S. Sawant Patricia Kovatch Benjamin Glicksberg Alexander Charney Girish N. Nadkarni Eric Oermann |
author_sort | Faris F. Gulamali |
collection | DOAJ |
description | Abstract Sample size estimation is a crucial step in experimental design but is understudied in the context of deep learning. Currently, estimating the quantity of labeled data needed to train a classifier to a desired performance, is largely based on prior experience with similar models and problems or on untested heuristics. In many supervised machine learning applications, data labeling can be expensive and time-consuming and would benefit from a more rigorous means of estimating labeling requirements. Here, we study the problem of estimating the minimum sample size of labeled training data necessary for training computer vision models as an exemplar for other deep learning problems. We consider the problem of identifying the minimal number of labeled data points to achieve a generalizable representation of the data, a minimum converging sample (MCS). We use autoencoder loss to estimate the MCS for fully connected neural network classifiers. At sample sizes smaller than the MCS estimate, fully connected networks fail to distinguish classes, and at sample sizes above the MCS estimate, generalizability strongly correlates with the loss function of the autoencoder. We provide an easily accessible, code-free, and dataset-agnostic tool to estimate sample sizes for fully connected networks. Taken together, our findings suggest that MCS and convergence estimation are promising methods to guide sample size estimates for data collection and labeling prior to training deep learning models in computer vision. |
first_indexed | 2024-03-11T13:47:12Z |
format | Article |
id | doaj.art-58170488a06d42cbb95fa02c45949816 |
institution | Directory Open Access Journal |
issn | 2398-6352 |
language | English |
last_indexed | 2024-03-11T13:47:12Z |
publishDate | 2022-12-01 |
publisher | Nature Portfolio |
record_format | Article |
series | npj Digital Medicine |
spelling | doaj.art-58170488a06d42cbb95fa02c459498162023-11-02T10:12:11ZengNature Portfolionpj Digital Medicine2398-63522022-12-01511810.1038/s41746-022-00728-0Autoencoders for sample size estimation for fully connected neural network classifiersFaris F. Gulamali0Ashwin S. Sawant1Patricia Kovatch2Benjamin Glicksberg3Alexander Charney4Girish N. Nadkarni5Eric Oermann6Icahn School of MedicineIcahn School of MedicineIcahn School of MedicineIcahn School of MedicineIcahn School of MedicineIcahn School of MedicineNew York UniversityAbstract Sample size estimation is a crucial step in experimental design but is understudied in the context of deep learning. Currently, estimating the quantity of labeled data needed to train a classifier to a desired performance, is largely based on prior experience with similar models and problems or on untested heuristics. In many supervised machine learning applications, data labeling can be expensive and time-consuming and would benefit from a more rigorous means of estimating labeling requirements. Here, we study the problem of estimating the minimum sample size of labeled training data necessary for training computer vision models as an exemplar for other deep learning problems. We consider the problem of identifying the minimal number of labeled data points to achieve a generalizable representation of the data, a minimum converging sample (MCS). We use autoencoder loss to estimate the MCS for fully connected neural network classifiers. At sample sizes smaller than the MCS estimate, fully connected networks fail to distinguish classes, and at sample sizes above the MCS estimate, generalizability strongly correlates with the loss function of the autoencoder. We provide an easily accessible, code-free, and dataset-agnostic tool to estimate sample sizes for fully connected networks. Taken together, our findings suggest that MCS and convergence estimation are promising methods to guide sample size estimates for data collection and labeling prior to training deep learning models in computer vision.https://doi.org/10.1038/s41746-022-00728-0 |
spellingShingle | Faris F. Gulamali Ashwin S. Sawant Patricia Kovatch Benjamin Glicksberg Alexander Charney Girish N. Nadkarni Eric Oermann Autoencoders for sample size estimation for fully connected neural network classifiers npj Digital Medicine |
title | Autoencoders for sample size estimation for fully connected neural network classifiers |
title_full | Autoencoders for sample size estimation for fully connected neural network classifiers |
title_fullStr | Autoencoders for sample size estimation for fully connected neural network classifiers |
title_full_unstemmed | Autoencoders for sample size estimation for fully connected neural network classifiers |
title_short | Autoencoders for sample size estimation for fully connected neural network classifiers |
title_sort | autoencoders for sample size estimation for fully connected neural network classifiers |
url | https://doi.org/10.1038/s41746-022-00728-0 |
work_keys_str_mv | AT farisfgulamali autoencodersforsamplesizeestimationforfullyconnectedneuralnetworkclassifiers AT ashwinssawant autoencodersforsamplesizeestimationforfullyconnectedneuralnetworkclassifiers AT patriciakovatch autoencodersforsamplesizeestimationforfullyconnectedneuralnetworkclassifiers AT benjaminglicksberg autoencodersforsamplesizeestimationforfullyconnectedneuralnetworkclassifiers AT alexandercharney autoencodersforsamplesizeestimationforfullyconnectedneuralnetworkclassifiers AT girishnnadkarni autoencodersforsamplesizeestimationforfullyconnectedneuralnetworkclassifiers AT ericoermann autoencodersforsamplesizeestimationforfullyconnectedneuralnetworkclassifiers |