This GitHub repository is describing parts of my master-thesis with the title "Analysis of existing methods for the interpretation of the decision process of machine learning methods regarding their suitablity for automotive applications"
For a detailed overview of the most used and most cited methods, please refer to the following publication: here
Methods of interpretable machine learning can help you to understand how models behave and make certain predictions.
Methods of interpretable machine learning can be very helpful for:
- Identifying bugs in your model and optimizing
- Getting insights how certain predictions were made
- Getting insights how models behave globally (Feature interaction)
- Detecting bias in traning data
- Verfiying legal requierements for model
- Verfiying confidence of model and prediction
- Creating an interface between Humans and AI
There are several ways to make machine leraning models and their predictions more transparent and interpretable. First of all, a distinction is made between global or local and model-specific or model-agnostic explanations of models. Following table gives a brief summary about some methods that I am analyzing in my thesis:
Method | Use | Scope |
---|---|---|
LIME* | model-agnostic | Local |
DeepLIFT | model-specific | Local |
Class Activation Maps | model-specific | Local |
Partial Dependence | model-agnostic | Global |
ELI5!** | model-agnostic | Global |
ICE*** | model-agnostic | Global |
You can find an implementation of LIME and Class Activation Maps in this Notebook other implemetations will follow.
* LIME: Local interpretable model-agnostic explanations
** ELI5!: Explain Lime I am 5!
*** ICE: Individual Conditional Expecation