Explainable machine learning explained

An AI classified a husky as a wolf because of the snow in the background. This is why explainability is crucial for building reliable models.

#1about 2 minutes

The growing importance of explainable AI in modern systems

Machine learning has become widespread, creating a critical need to understand how models make decisions beyond simple accuracy metrics.

#2about 4 minutes

Why regulated industries like medtech and fintech require explainability

In fields like medicine and finance, regulatory compliance and user trust make it mandatory to explain how AI models arrive at their conclusions.

#3about 3 minutes

Identifying the key stakeholders who need model explanations

Explainability is crucial for various roles, including domain experts like doctors, regulatory agencies, business leaders, data scientists, and end-users.

#4about 4 minutes

Fundamental approaches for explaining AI model behavior

Models can be explained through various methods such as mathematical formulas, visual charts, local examples, simplification, and analyzing feature relevance.

#5about 5 minutes

Learning from classic machine learning model failures

Examining famous failures, like the husky vs. wolf classification and the Tay chatbot, reveals how models can learn incorrect patterns from biased data.

#6about 5 minutes

Differentiating between white-box and black-box models

White-box models like decision trees are inherently transparent, whereas black-box models like neural networks require special techniques to interpret their internal workings.

#7about 7 minutes

Improving model performance with data-centric feature engineering

A data-centric approach, demonstrated with the Titanic dataset, shows how creating new features from existing data can significantly boost model accuracy.

#8about 4 minutes

Exploring inherently interpretable white-box models

Models such as logistic regression, k-means, decision trees, and SVMs are considered explainable by design due to their transparent decision-making processes.

#9about 5 minutes

Using methods like LIME and SHAP to explain black-box models

Techniques like Partial Dependence Plots (PDP), LIME, and SHAP are used to understand the influence of features on the predictions of complex black-box models.

#10about 3 minutes

Visualizing deep learning decisions in images with Grad-CAM

Grad-CAM (Gradient-weighted Class Activation Mapping) creates heatmaps to highlight which parts of an image were most influential for a deep neural network's classification.

#11about 3 minutes

Understanding security risks from adversarial attacks on models

Adversarial attacks demonstrate how small, often imperceptible, changes to input data can cause machine learning models to make completely wrong predictions.

From learning to earning

Jobs that call for the skills explored in this talk.