explainable machine learning
explainable machine learning
Explainable Machine Learning Open Course
1 Introduction
Regarding tanks : When identifying real and fake tanks in pictures, AI mistakenly used weather information as a criterion for identifying tanks.
in conclusion:
- The training set and test set should come from a distribution
- Neural Network -> "Black Box"
Machine learning is mostly statistical learning, that is, fitting to data .
In low dimension, classification is to use a curve to separate the samples, and regression is to use a curve to fit the sample data distribution. In high dimensions, the whole process is a "black box".
2. Why study explainable machine learning
3. Machine learning algorithms with good interpretability
- KNN
- logistic regression
- linear regression
- decision tree
- Naive Bayes
Interpretability Analysis of Traditional Machine Learning Algorithms
- Visualization that comes with the algorithm
- The feature weights that come with the algorithm
- Permutation Importance Permutation importance (judgment feature is disrupted to judge whether the feature is important)
- PDP diagram, ICE diagram
- Shapley 值
- Lime
4. Deep learning is poorly interpretable
Interpretability Analysis of Convolutional Neural Networks
- Visualize convolution kernel, feature map
- Occlusion, scaling, translation, rotation
- Find the original image pixel or small image that can activate a neuron
- Visualization based on class activation heat map (CAM)
- Semantic Coding Dimensionality Reduction Visualization
- The original image input by semantic coding backwards
- Generate images that meet certain requirements (a category with the highest probability of prediction)
5. Summary
Some references:
Interpretable Analysis, Significance Analysis Code Practice
Pytorch-cnn-visualizations
Summary Paper