Opening the Black Box: Bootstrapping Sensitivity Measures in Neural Networks for Interpretable Machine Learning

Rocca, Michele La; Perna, Cira

doi:10.3390/stats5020026

Cited by 5 publications

(5 citation statements)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, significant steps still need to be taken, as the inherent “black box” nature of deep learning poses challenges in interpreting deep learning models. Anticipated efforts in enhancing the interpretability of the model are expected to refine its overall effectiveness and integration into climate research methodologies (Guidotti et al., 2018; La Rocca & Perna, 2022; Savage, 2022).…”

Section: Discussionmentioning

confidence: 99%

Physically‐Informed Super‐Resolution Downscaling of Antarctic Surface Melt

de Roda Husman,

Hu,

van Tiggelen

et al. 2024

J Adv Model Earth Syst

View full text Add to dashboard Cite

Because Antarctic surface melt is mostly driven by local processes, its simulation necessitates high‐resolution regional climate models (RCMs). However, the current horizontal resolution of RCMs (≈25–30 km) is inadequate for capturing small‐scale melt processes. To address this limitation, we present SUPREME (SUPer‐REsolution‐based Melt Estimation over Antarctica), a deep learning method to downscale surface melt to 5.5 km resolution using a physically‐informed super‐resolution model. The physical information integrated into the model originates from observations tied to surface melt, specifically remote sensing‐derived albedo and elevation. These remote sensing data, in addition to a Regional Atmospheric Climate Model (RACMO) run at 27 km resolution, account for the diverse drivers of surface melt across Antarctica, facilitating effective generalization beyond the training region of the Antarctic Peninsula. A comparison of SUPREME with a dynamically downscaled RACMO run at 5.5 km over the Antarctic Peninsula shows high accuracy, with average yearly RMSE and bias of 5.5 mm w.e. yr−1 and 4.5 mm w.e. yr−1, respectively. Validation at five automatic weather stations reveals SUPREME's marked improvement with substantially lower average RMSE (81 mm w.e.) compared to RACMO 27 km (129 mm w.e.). Beyond the training region, SUPREME aligns more closely with remote sensing products associated with surface melt than super‐resolution models lacking physical constraints. While further validation of SUPREME is needed, our study highlights the potential of super‐resolution techniques with physical constraints for high‐resolution surface melt monitoring in Antarctica, providing insights into the impacts of localized melting on processes affecting ice shelf integrity such as hydrofracturing.

show abstract

Section: Discussionmentioning

confidence: 99%

Physically‐Informed Super‐Resolution Downscaling of Antarctic Surface Melt

de Roda Husman,

Hu,

van Tiggelen

et al. 2024

J Adv Model Earth Syst

View full text Add to dashboard Cite

show abstract

“…First, neural networks, as used here, have a reputation for being a black box algorithm and thus having a decision process that is hard to understand. Still, there are recent developments to make them more transparent (Guidotti et al., 2019; Rocca & Perna, 2022; Savage, 2022). Second, training a neural network while relying on a GPU creates several sources of randomness, and reproducibility depends on using the exact same settings as the authors of a framework (Feng & Hao, 2020; Scardapane & Wang, 2017).…”

Section: Discussionmentioning

confidence: 99%

Deep Learning Regional Climate Model Emulators: A Comparison of Two Downscaling Training Frameworks

Meer

Husman

Lhermitte

2023

J Adv Model Earth Syst

View full text Add to dashboard Cite

Regional climate models (RCMs) have a high computational cost due to their higher spatial resolution compared to global climate models (GCMs). Therefore, various downscaling approaches have been developed as a surrogate for the dynamical downscaling of GCMs. This study assesses the potential of using a cost‐efficient machine learning alternative to dynamical downscaling by using the example case study of emulating surface mass balance (SMB) over the Antarctic Peninsula. More specifically, we determine the impact of the training framework by comparing two training scenarios: (a) a perfect and (b) an imperfect model framework. In the perfect model framework, the RCM‐emulator learns only the downscaling function; therefore, it was trained with upscaled RCM (UPRCM) features at GCM resolution. This emulator accurately reproduced SMB when evaluated on UPRCM, but its predictions on GCM data conserved RCM‐GCM inconsistencies and led to underestimation. In the imperfect model framework, the RCM‐emulator was trained with GCM features and downscaled the GCM while exposed to RCM‐GCM inconsistencies. This emulator predicted SMB close to the truth, showing it learned the underlying inconsistencies and dynamics. Our results suggest that a deep learning RCM‐emulator can learn the proper GCM to RCM downscaling function while working directly with GCM data. Furthermore, the RCM‐emulator presents a significant computational gain compared to an RCM simulation. We conclude that machine learning emulators can be applied to produce fast and fine‐scaled predictions of RCM simulations from GCM data.

show abstract

Section: Limitations Of Machine Learningmentioning

confidence: 99%

Deep Learning Regional Climate Model Emulators: a comparison of two downscaling training frameworks

Meer

Husman

Lhermitte

2022

Preprint

View full text Add to dashboard Cite

Regional climate models (RCMs) have a high computational cost due to their higher spatial resolution compared to global climate models (GCMs). Therefore, various downscaling approaches have been developed as a surrogate for the dynamical downscaling of GCMs. This study assesses the potential of using a cost-efficient machine learning alternative to dynamical downscaling by using the example case study of emulating surface mass balance (SMB) over the Antarctic Peninsula. More specifically, we determine the impact of the training framework by comparing two training scenarios: (1) a perfect and (2) an imperfect model framework. In the perfect model framework, the RCM-emulator learns only the downscaling function, similar to classical super-resolution approaches; therefore, it was trained with upscaled RCM features at GCM resolution. This emulator accurately reproduced SMB when evaluated on upscaled RCM features (mean RMSE of 0.27 mm w.e./day), but its predictions on GCM data conserved RCM-GCM inconsistencies and led to underestimation. In the imperfect model framework, the RCM-emulator was trained with GCM features and downscaled the GCM while exposed to RCM-GCM inconsistencies. This emulator predicted SMB close to the truth, showing it learned the underlying inconsistencies and dynamics. Our results suggest that a deep learning RCM-emulator can learn the proper GCM to RCM downscaling function while working directly with GCM data. Furthermore, the RCM-emulator presents a significant computational gain compared to an RCM simulation. We conclude that machine learning emulators can be applied to produce fast and fine-scaled predictions of RCM simulations from GCM data.

show abstract

Opening the Black Box: Bootstrapping Sensitivity Measures in Neural Networks for Interpretable Machine Learning

Cited by 5 publications

References 28 publications

Physically‐Informed Super‐Resolution Downscaling of Antarctic Surface Melt

Physically‐Informed Super‐Resolution Downscaling of Antarctic Surface Melt

Deep Learning Regional Climate Model Emulators: A Comparison of Two Downscaling Training Frameworks

Deep Learning Regional Climate Model Emulators: a comparison of two downscaling training frameworks

Contact Info

Product

Resources

About