Don't Lie to Me! Robust and Efficient Explainability with Verified Perturbation Analysis
Authors:
Thomas Fel,
Melanie Ducoffe,
David Vigouroux,
Remi Cadene,
Mikael Capelle,
Claire Nicodeme,
Thomas Serre
Abstract:
A variety of methods have been proposed to try to explain how deep neural networks make their decisions. Key to those approaches is the need to sample the pixel space efficiently in order to derive importance maps. However, it has been shown that the sampling methods used to date introduce biases and other artifacts, leading to inaccurate estimates of the importance of individual pixels and severe…
▽ More
A variety of methods have been proposed to try to explain how deep neural networks make their decisions. Key to those approaches is the need to sample the pixel space efficiently in order to derive importance maps. However, it has been shown that the sampling methods used to date introduce biases and other artifacts, leading to inaccurate estimates of the importance of individual pixels and severely limit the reliability of current explainability methods. Unfortunately, the alternative -- to exhaustively sample the image space is computationally prohibitive. In this paper, we introduce EVA (Explaining using Verified perturbation Analysis) -- the first explainability method guarantee to have an exhaustive exploration of a perturbation space. Specifically, we leverage the beneficial properties of verified perturbation analysis -- time efficiency, tractability and guaranteed complete coverage of a manifold -- to efficiently characterize the input variables that are most likely to drive the model decision. We evaluate the approach systematically and demonstrate state-of-the-art results on multiple benchmarks.
△ Less
Submitted 18 March, 2023; v1 submitted 15 February, 2022;
originally announced February 2022.
Schedule Earth Observation satellites with Deep Reinforcement Learning
Authors:
Adrien Hadj-Salah,
Rémi Verdier,
Clément Caron,
Mathieu Picard,
Mikaël Capelle
Abstract:
Optical Earth observation satellites acquire images worldwide , covering up to several million square kilometers every day. The complexity of scheduling acquisitions for such systems increases exponentially when considering the interoperabil-ity of several satellite constellations together with the uncertainties from weather forecasts. In order to deliver valid images to customers as fast as possi…
▽ More
Optical Earth observation satellites acquire images worldwide , covering up to several million square kilometers every day. The complexity of scheduling acquisitions for such systems increases exponentially when considering the interoperabil-ity of several satellite constellations together with the uncertainties from weather forecasts. In order to deliver valid images to customers as fast as possible, it is crucial to acquire cloud-free images. Depending on weather forecasts, up to 50% of images acquired by operational satellites can be trashed due to excessive cloud covers, showing there is room for improvement. We propose an acquisition scheduling approach based on Deep Reinforcement Learning and experiment on a simplified environment. We find that it challenges classical methods relying on human-expert heuristic.
△ Less
Submitted 12 November, 2019;
originally announced November 2019.
