Technical Barriers to the Adoption of Post-hoc Explanation Methods for Black Box AI models

Abstract We examine two key technical barriers to the adoption of post hoc methods for explaining the outputs of black box AI models: their lack of robustness, and the difficulty in assessing explanation fidelity.
  • Eunjin Lee (IBM UK)
  • Harrison Taylor (Cardiff)
  • Liam Hiley (Cardiff)
  • Richard Tomsett (IBM UK)
Date Apr-2021
Venue AISB 2021 Symposium, Overcoming Opacity in Machine Learning