A Learning Paradigm for Interpretable Gradients
Authors:
Felipe Torres Figueroa,
Hanwei Zhang,
Ronan Sicre,
Yannis Avrithis,
Stephane Ayache
Abstract:
This paper studies interpretability of convolutional networks by means of saliency maps. Most approaches based on Class Activation Maps (CAM) combine information from fully connected layers and gradient through variants of backpropagation. However, it is well understood that gradients are noisy and alternatives like guided backpropagation have been proposed to obtain better visualization at infere…
▽ More
This paper studies interpretability of convolutional networks by means of saliency maps. Most approaches based on Class Activation Maps (CAM) combine information from fully connected layers and gradient through variants of backpropagation. However, it is well understood that gradients are noisy and alternatives like guided backpropagation have been proposed to obtain better visualization at inference. In this work, we present a novel training approach to improve the quality of gradients for interpretability. In particular, we introduce a regularization loss such that the gradient with respect to the input image obtained by standard backpropagation is similar to the gradient obtained by guided backpropagation. We find that the resulting gradient is qualitatively less noisy and improves quantitatively the interpretability properties of different networks, using several interpretability methods.
△ Less
Submitted 23 April, 2024;
originally announced April 2024.
Energy Loss Due to Defect Formation from $^{206}$Pb Recoils in SuperCDMS Germanium Detectors
Authors:
Robert Agnese,
Taylor Aralis,
Tsuguo Aramaki,
Isaac Arnquist,
Elham Azadbakht,
William Baker,
Samir Banik,
D'Ann Barker,
Dan Bauer,
Thomas Binder,
Michael Bowles,
Paul Brink,
Ray Bunker,
Blas Cabrera,
Robert Calkins,
Concetta Cartaro,
David Cerdeno,
Yen-Yung Chang,
Jodi Cooley,
Brett Cornell,
Priscilla Cushman,
Philippe Di Stefano,
Todd Doughty,
Eleanor Fascione,
Tali Figueroa
, et al. (72 additional authors not shown)
Abstract:
The Super Cryogenic Dark Matter Search experiment (SuperCDMS) at the Soudan Underground Laboratory studied energy loss associated with Frenkel defect formation in germanium crystals at mK temperatures using in situ $^{210}$Pb sources. We examine the spectrum of $^{206}$Pb nuclear recoils near its expected 103 keV endpoint energy and determine an energy loss of $\left(6.08\pm0.18\right)$ %, which w…
▽ More
The Super Cryogenic Dark Matter Search experiment (SuperCDMS) at the Soudan Underground Laboratory studied energy loss associated with Frenkel defect formation in germanium crystals at mK temperatures using in situ $^{210}$Pb sources. We examine the spectrum of $^{206}$Pb nuclear recoils near its expected 103 keV endpoint energy and determine an energy loss of $\left(6.08\pm0.18\right)$ %, which we attribute to defect formation. From this result and using TRIM simulations, we extract the first experimentally determined average displacement threshold energy of $\left(19.7^{+0.6}_{-0.5}\right)$ eV for germanium. This has implications for the analysis thresholds of future germanium-based dark matter searches.
△ Less
Submitted 16 April, 2019; v1 submitted 24 May, 2018;
originally announced May 2018.
