@incollection{fdi:010090489,
  title = {{O}ut-of-distribution detection is not all you need},
  author = {{G}uerin, {J}oris and {D}elmas, {K}. and {F}erreira, {R}. and {G}uiochet, {J}.},
  editor = {},
  language = {{ENG}},
  abstract = {{T}he usage of deep neural networks in safety-critical systems is limited by our ability to guarantee their correct behavior. {R}untime monitors are components aiming to identify unsafe predictions and discard them before they can lead to catastrophic consequences. {S}everal recent works on runtime monitoring have focused on out-of-distribution ({OOD}) detection, i.e., identifying inputs that are different from the training data. {I}n this work, we argue that {OOD} detection is not a well-suited framework to design efficient runtime monitors and that it is more relevant to evaluate monitors based on their ability to discard incorrect predictions. {W}e call this setting out-of-model-scope detection and discuss the conceptual differences with {OOD}. {W}e also conduct extensive experiments on popular datasets from the literature to show that studying monitors in the {OOD} setting can be misleading: 1. very good {OOD} results can give a false impression of safety, 2. comparison under the {OOD} setting does not allow identifying the best monitor to detect errors. {F}inally, we also show that removing erroneous training data samples helps to train better monitors.},
  keywords = {},
  booktitle = {{P}roceedings of the {T}hirty-{S}eventh {AAAI} {C}onference on {A}rtificial {I}ntelligence and {T}hirty-{F}ifth {C}onference on {I}nnovative {A}pplications of {A}rtificial {I}ntelligence and {T}hirteenth {S}ymposium on {E}ducational {A}dvances in {A}rtificial {I}ntelligence},
  numero = {},
  pages = {14829--14837},
  address = {{W}ashington},
  publisher = {{AAAI} {P}ress},
  series = {},
  year = {2023},
  DOI = {10.1609/aaai.v37i12.26732},
  ISBN = {978-1-57735-880-0},
  URL = {https://www.documentation.ird.fr/hor/fdi:010090489},
}