Skip to main content

Advanced topis part 2

  • Most of AI and ML focuses on full automation
  • Bigger and bigger datasets/models
  • Good for benchmarking, less good for actual applicaiton
  • Datasets are saturated with highly specific models
  • Models specialise so much they cannot adapt

Need to keep humans in the loop​

  • Interactive sense-making: Visualisation + Analysis + Training
  • Explainable AI: Explanation + Analysis + Training
  • Adversarial training with humans in the loop: Generative Adversarial Networks + Training
  • Active learning: Sampling + Training
  • Meta Learning: Human-guided training

Interpretable Machine Learning​

Interpretability brings advantages:

  • Human expert can double check a result
  • Human user can therefore take liability
  • A ML/AI developer can debug the inference process Does bring drawbacks:
  • Additional drawbacks
  • Comes at performance cost

Taxonomy of interpretability​

  • Intrinsic vs post hoc
  • Model-specific vs model-agnostic
  • Local vs Global

Model-agnostic methods​

....

Global Surrogate Model​

Use an interpretable model that is trained to approximate the predictions of a black box model

Local Surrogate Model​

  • Local surrogate models are used to explain individual predictions of black box machine learning models

Explaining with examples​

...

Counterfactual Explanation​

Prototypes​

  • Finding instances which are representative of a class
  • A criticism is an instance that is not well covered by prototypes

Active Learning​

Motivation​

  • Traditional supervised learning
    • Labelled training examples are randomly sampled from the training set and fed into the ML algorithm
  • Active Learning
    • The ML algorithm starts from scratch and asks the user to label specific data points

Most Popular Approach: Uncertainty sampling .....