Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Overall discussion for Tip 8 #249

Open
SiminaB opened this issue Oct 7, 2020 · 1 comment
Open

Overall discussion for Tip 8 #249

SiminaB opened this issue Oct 7, 2020 · 1 comment

Comments

@SiminaB
Copy link
Collaborator

SiminaB commented Oct 7, 2020

This is to discuss outstanding issues for Tip 8: Your DL models can be more transparent.
https://github.com/Benjamin-Lee/deep-rules/blob/master/content/10.blackbox.md

@SiminaB
Copy link
Collaborator Author

SiminaB commented Oct 7, 2020

  • I would suggest changing the title to "Aim to have interpretable DL model" to make it more in line with the other tips.

  • Some references should be included in the last paragraph discussing model interpretation:

Model interpretation is an open, active area of research. It is becoming more feasible to interpret models with many parameters and non-linear relationships, but in many cases simpler models remain substantially easier to interpret than more complex ones. When deciding on a machine learning approach and model architecture, consider an interpretability versus accuracy tradeoff. A challenge in considering this tradeoff is that the extent to which one trades interpretability for accuracy depends on the problem itself. When the features provided to the model are already highly relevant to the task at hand, a simpler, interpretable model that gives up only a little performance when compared to a very complex one more useful in many settings. On the other hand, if features must be combined in complex ways to be meaningful for the task, the performance difference of a model capable of capturing that structure may outweigh the interpretability costs. An appropriate choice can only be made after careful consideration, which often includes estimating the performance of a simple, linear model that serves as a baseline. In cases where models are learned from high-throughput datasets, a small subset of features in the dataset may be strongly correlated with the complex combination of the larger feature set defined from the deep learning model. In this case, this more limited number of features can themselves be used in the subsequent simplified model to further enhance interpretability of the model. This feature reduction can be essential to defining biomarker panels that enable clinical applications.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant