Week 11 - Explainability

Question 1

Why do de need reasons for interpretability in ML models. Name 4 reasons

Answer 1

1. Preventing Biases
2. Verifying decisions
3. Trust and acceptance
4. Debugging

Question 2

What is ‘explaining predictions’ in terms of ml

Answer 2

By explaining predictions we mean presenting textual or visual artifacts that provide a qualitative understanding of the relationship between the instance’s components and the model’s prediction

Question 3

What are interpretable models

Answer 3

Some models are inherently explainable. This allows for exploring why a model made a certain decision

Question 4

What are black box models

Answer 4

Some ml are considered black box. it is difficult to investigate and understand why a certain ml decision has been reaches

Question 5

Give 3 example models that are not inherently explainable

Answer 5

• Neural networks
• Random forest
• SVM kernels

Question 6

What are model-specific approaches for interpretation

Answer 6

model specific interpretation tools are limited to specific model classes.

Question 7

What are model agnostic approaches

Answer 7

Model-agnostic tools can be used on any mahine-learning model and are applied after the model has been trained. These agnositc moethods usually work by analyszing feature input and output pairs.

Question 8

What are the 3 interpretable model properties

Answer 8

1. Linear
2. Monotonicity: the relationship between a feature and the target outcome always goes in the same direction over the entire range of the feature
3. Interaction: automatically includes interactions between features to predict the target outcome.

Question 9

How can we interpret a linear regression model?

Answer 9

1. Numerical feauture.
2. Categorical feature
3. Intercept

Question 10

What are 3 model-agnostic methods of interpretability

Answer 10

1. Model flexibility
2. Explanation flexibility
3. Representation flexibility

Question 11

What does LIME stand for

Answer 11

Local Interpretable Model-Agnostic Explanations

Question 12

What are surrogate models

Answer 12

Surrogate models are trained to approximate the predictions of the underlying black box model. Instead of training a global surrogate model, LIME focuses on training a local surrogate models to explain individual predictions. A surrogate model in machine learning is an approximate model used to mimic the behavior of a more complex or computationally expensive model

Question 13

What is Local fidelity

Answer 13

Fidelity shows how faithful the interpretation is to the while model. With non-linear models, it might be impossible to find an explanation that is faithfil unless it is the complete desciption of the model itself.

Question 14

What is the difference between local fidelity and global fidelity

Answer 14

Local Fidelity: The accuracy of a surrogate model in approximating the true model’s behavior in a specific, localized region of the input space.

Global Fidelity: The accuracy of a surrogate model in approximating the true model’s behavior across the entire input space.

Question 15

The idea of LIME is to come up with a balanced trade-off between ____ and ____

Answer 15

interpretability and local fidelity

Question 16

What is the LIME approach. Give steps

Answer 16

• we have a black box model where you can input data points and get predictions of the model.
• We can probe the box as often as you want to understand why the machine learning model made a certain prediction. LIME tests what happens to the predictions when you give variations of your data(perturbation) into the machine learning model.
• LIME generates a new dataset
• LIME then trains an interpretable model

Question 17

What are perturbed samples.

Answer 17

Data points that have been intentionally modified or slightly altered from their original values to explore the sensitivity of a model or to enhance robustness during training.

Question 18

What are the steps in LIME for interpreting image classification

Answer 18

1. Modify the image by enabling and disambling some of the interpeatble components
2. For each of the generated images the probability of a tree frog’s preseness on it is obtained and used to create a labeled data-set.
3. Next we train a lienar model weighted on the created data-set
4. The components with the highest weights are presented