11.2 Method: SpRay

An international cooperation investigated recent ML successes in order to find whether or not the models deliver reliably for the problems they are trained for (Lapuschkin et al. 2019)

Basics of SpRay:

  • Identifies and quantifies decision-making behaviors
  • Finds undesirable decisions in vast data sets
  • Can be used to improve
    • model or
    • dataset

Lets examine a model which classifies horses

Figure from (Lapuschkin et al. 2019)

SpRay identifies 4 different prediction strategies for classifying images as “horse”

Detected prediction strategies:

  • b \(\implies\) detect a horse (and rider)
  • c \(\implies\) detect a source tag in portrait oriented images
  • d \(\implies\) detect wooden hurdles and other contextual elements of horseback riding
  • e \(\implies\) detect a source tag in landscape-oriented images

References

Lapuschkin, Sebastian, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. 2019. “Unmasking Clever Hans Predictors and Assessing What Machines Really Learn.” Nature Communications 10 (1): 1–8.