Grant: Discovery Grants Program - Individual
The University of British Columbia
$40,000.00
Beyond the confusion matrix: A framework to compare neural network representations trained on biomedical data
FY 2024-25
Summary
Machine learning (a type of artificial intelligence) is starting to make a big impact in healthcare. In particular, deep learning, which uses artificial neural networks (ANNs), is being used to help make clinical decisions, such as classifying whether a medical scan shows an abnormality. Many variations of ANNs are being published in the literature with claims of being better than others for certain classification tasks, but often the performance numbers, such as accuracy, only span a small range (< 3%) between the presented methods. In such cases, it is difficult to know whether the authors' proposed ANN variant is truly learning something new and relevant to the task from the data used to train it, or whether the differences are due to other, potentially more random, factors.
ANNs process information in layers, with the first and last layers being the input and output, respectively. The information captured in the intermediate layers, called the "hidden" layers, can reveal how data is being represented internally by the ANN. For example, for an ANN used to analyze images, examining its internal representation can reveal which part(s) of an image the ANN is using to make a decision. To determine whether two ANNs are learning similar patterns from a given data set, we can try to measure how different their internal representations are. Many representational similarity measures, which are mathematical formulas for comparing representations, have been published, but there has been very little research to systematically compare similarity measures, particularly for biomedical data. Biomedical data sets are typically much smaller but more complex than the ones used in machine learning papers. There is very little knowledge on which representational similarity measures are useful for which combinations of AANs and data types.
Our proposed research program will develop a new framework to measure the differences in the representations learned by ANNs used for classifying biomedical data. We will develop this framework using a large variety of completed, ongoing, and future studies with different types of biomedical data. The knowledge gained from our proposed research will allow scientists working with biomedical data to select the appropriate representational similarity measures to use for their given experiment setup, and to have a deeper understanding of whether any observed differences in performance between AANs are due to differences in what the models are learning. This type of knowledge is critical for driving further improvements in AANs.
Program
Discovery Grants Program - Individual
Recipient
Location
British Columbia, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
RGPIN-2024-04718
Name nom
Tam, Roger
Department département
School of Biomedical Engineering
Organization id
2
Institution établissement
The University of British Columbia
Province en
British Columbia
Province fr
Colombie-Britannique
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2024
Award amount
40000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1510
Committee name en
Electrical and Computer Engineering
Committee name fr
Génie électrique et informatique
Area of application code
1102
Area of application group en
Health, education and social services
Area of application group fr
Santé publique, éducation et services sociaux
Area of application en
Biomedical engineering
Area of application fr
Génie biomédical
Research subject code
2800
Research subject group en
Artificial intelligence
Research subject group fr
Intelligence artificielle
Research subject en
Artificial intelligence (computer vision, use2603)
Research subject fr
Intelligence artificielle (vision artificielle, utiliser 2603)
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Beyond the confusion matrix: A framework to compare neural network representations trained on biomedical data
Keywords
biomedical signal analysis; machine learning; classification; neural networks; deep learning representations; similarity measures
Application summary
Machine learning (a type of artificial intelligence) is starting to make a big impact in healthcare. In particular, deep learning, which uses artificial neural networks (ANNs), is being used to help make clinical decisions, such as classifying whether a medical scan shows an abnormality. Many variations of ANNs are being published in the literature with claims of being better than others for certain classification tasks, but often the performance numbers, such as accuracy, only span a small range (< 3%) between the presented methods. In such cases, it is difficult to know whether the authors' proposed ANN variant is truly learning something new and relevant to the task from the data used to train it, or whether the differences are due to other, potentially more random, factors.
ANNs process information in layers, with the first and last layers being the input and output, respectively. The information captured in the intermediate layers, called the "hidden" layers, can reveal how data is being represented internally by the ANN. For example, for an ANN used to analyze images, examining its internal representation can reveal which part(s) of an image the ANN is using to make a decision. To determine whether two ANNs are learning similar patterns from a given data set, we can try to measure how different their internal representations are. Many representational similarity measures, which are mathematical formulas for comparing representations, have been published, but there has been very little research to systematically compare similarity measures, particularly for biomedical data. Biomedical data sets are typically much smaller but more complex than the ones used in machine learning papers. There is very little knowledge on which representational similarity measures are useful for which combinations of AANs and data types.
Our proposed research program will develop a new framework to measure the differences in the representations learned by ANNs used for classifying biomedical data. We will develop this framework using a large variety of completed, ongoing, and future studies with different types of biomedical data. The knowledge gained from our proposed research will allow scientists working with biomedical data to select the appropriate representational similarity measures to use for their given experiment setup, and to have a deeper understanding of whether any observed differences in performance between AANs are due to differences in what the models are learning. This type of knowledge is critical for driving further improvements in AANs.