Deep Learning Algorithm Performance
Principal Investigator
Name
Gregory Sorensen
Degrees
MD
Institution
N/A
Position Title
researcher
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-152
Initial CDAS Request Approval
Aug 18, 2015
Title
Deep Learning Algorithm Performance
Summary
Deep learning algorithms have shown remarkable advances in many artificial intelligence domains. However, in many situations, the problem space has been purposely constrained, and this might lead to estimations of overperformance of deep learning algorithms. Therefore, one goal of this project is to quantitatively assess the performance of deep learning approaches as the problem space is increasingly enlarged (such as a whole CT scan of the body).
An attractive feature of many currently popular deep learning algorithms is their ability to perform well with essentially no training other than exposure to training dataset. That is, unlike older methods of pattern recognition or machine learning, the algorithms are not taught features that might be of interest (such as edges within an image). This property comes with a downside, however: if a given feature set is known to be highly indicative of a given classification outcome, there appears to be no a priori way to instruct the deep learning algorithms. Such a drawback becomes more evident in situations where relatively few training data are available. For example, certain morphometric features of cancers such as a pattern of calcification might be rare but highly indicative of diagnosis. Finding enough training cases might be impossible, and this could limit the overall value of deep learning approaches. This project therefore seeks to explore such potential performance limitations.
An attractive feature of many currently popular deep learning algorithms is their ability to perform well with essentially no training other than exposure to training dataset. That is, unlike older methods of pattern recognition or machine learning, the algorithms are not taught features that might be of interest (such as edges within an image). This property comes with a downside, however: if a given feature set is known to be highly indicative of a given classification outcome, there appears to be no a priori way to instruct the deep learning algorithms. Such a drawback becomes more evident in situations where relatively few training data are available. For example, certain morphometric features of cancers such as a pattern of calcification might be rare but highly indicative of diagnosis. Finding enough training cases might be impossible, and this could limit the overall value of deep learning approaches. This project therefore seeks to explore such potential performance limitations.
Aims
1. Explore the impact of the size of the problem space on deep learning performance.
2. Explore the impact of training size versus features on deep learning performance.