"Who experiences large model decay and why?" A Hierarchical Framework for Diagnosing Heterogeneous Performance Drift
Authors:
Harvineet Singh,
Fan Xia,
Alexej Gossmann,
Andrew Chuang,
Julian C. Hong,
Jean Feng
Abstract:
Machine learning (ML) models frequently experience performance degradation when deployed in new contexts. Such degradation is rarely uniform: some subgroups may suffer large performance decay while others may not. Understanding where and how large differences in performance arise is critical for designing targeted corrective actions that mitigate decay for the most affected subgroups while minimiz…
▽ More
Machine learning (ML) models frequently experience performance degradation when deployed in new contexts. Such degradation is rarely uniform: some subgroups may suffer large performance decay while others may not. Understanding where and how large differences in performance arise is critical for designing targeted corrective actions that mitigate decay for the most affected subgroups while minimizing any unintended effects. Current approaches do not provide such detailed insight, as they either (i) explain how average performance shifts arise or (ii) identify adversely affected subgroups without insight into how this occurred. To this end, we introduce a Subgroup-scanning Hierarchical Inference Framework for performance drifT (SHIFT). SHIFT first asks "Is there any subgroup with unacceptably large performance decay due to covariate/outcome shifts?" (Where?) and, if so, dives deeper to ask "Can we explain this using more detailed variable(subset)-specific shifts?" (How?). In real-world experiments, we find that SHIFT identifies interpretable subgroups affected by performance decay, and suggests targeted actions that effectively mitigate the decay.
△ Less
Submitted 31 May, 2025;
originally announced June 2025.
Morphological evidence for nanoflares heating warm loops in the solar corona
Authors:
Y. Bi,
J. J. Yang,
Y. Qin,
Z. P. Qiang,
J. C. Hong,
B. Yang,
Z. Xu,
H. Liu,
K. F. Ji
Abstract:
Nanoflares are impulsive energy releases by magnetic reconnection in the braided coronal magnetic field, which is a potential mechanism for heating the corona. However, there are still sporadic observations of the interchange of braiding structure segments and footpoints inside coronal loops, which is predicted to be the morphological evolution of the reconnecting magnetic bundles in the nanoflare…
▽ More
Nanoflares are impulsive energy releases by magnetic reconnection in the braided coronal magnetic field, which is a potential mechanism for heating the corona. However, there are still sporadic observations of the interchange of braiding structure segments and footpoints inside coronal loops, which is predicted to be the morphological evolution of the reconnecting magnetic bundles in the nanoflare picture. This work aims to detect the evolutions of the pairs of braiding strands within the apparent single coronal loops observed in Atmospheric Imaging Assembly (AIA) images. The loop strands are detected on two kinds of upsampled AIA 193 Å images, which are obtained by upscaling the Point Spread Function matched AIA images via Bicubic interpolation and are generated using a super-resolution convolutional neural network, respectively. The architecture of the network is designed to map the AIA images to unprecedentedly high spatial resolution coronal images taken by High-resolution Coronal Imager (Hi-C) during its brief flight. At times, pairs of separate strands that appear braided together later evolved into pairs of almost parallel strands with completely exchanged parts. These evolutions offer morphological evidence that magnetic reconnections between the braiding strands have taken place, which is further supported by the appearance of transient hot emissions containing significant high-temperature components (T > 5MK) at the footpoints of the braiding structures. The brief appearances of the two rearranging strands support that magnetic reconnections have occurred within what appears to be a single AIA loop.
△ Less
Submitted 18 September, 2023;
originally announced September 2023.
