Efficient Computation of Shap Explanation Scores for Neural Network Classifiers via Knowledge Compilation
Authors:
Leopoldo Bertossi,
Jorge E. Leon
Abstract:
The use of Shap scores has become widespread in Explainable AI. However, their computation is in general intractable, in particular when done with a black-box classifier, such as neural network. Recent research has unveiled classes of open-box Boolean Circuit classifiers for which Shap can be computed efficiently. We show how to transform binary neural networks into those circuits for efficient Sh…
▽ More
The use of Shap scores has become widespread in Explainable AI. However, their computation is in general intractable, in particular when done with a black-box classifier, such as neural network. Recent research has unveiled classes of open-box Boolean Circuit classifiers for which Shap can be computed efficiently. We show how to transform binary neural networks into those circuits for efficient Shap computation.We use logic-based knowledge compilation techniques. The performance gain is huge, as we show in the light of our experiments.
△ Less
Submitted 22 July, 2023; v1 submitted 11 March, 2023;
originally announced March 2023.
Fundamental effective temperature measurements for eclipsing binary stars -- III. SPIRou near-infrared spectroscopy and CHEOPS photometry of the benchmark G0V star EBLM J0113+31
Authors:
P. F. L. Maxted,
N. J. Miller,
S. Hoyer,
V. Adibekyan,
S. G. Sousa,
N. Billot,
A. Fortier,
A. E. Simon,
A. Collier Cameron,
M. I. Sawyne,
P. Gutermann,
A. H. M. J. Triaud,
J. Southworth,
Y. Alibert,
R. Alonso,
G. Anglada,
T. Bárczy,
D. Barrado y Navascues,
S. C. C. Barros,
W. Baumjohann,
M. Beck,
T. Beck,
W. Benz,
X. Bonfils,
A. Brandeker
, et al. (52 additional authors not shown)
Abstract:
EBLM J0113+31 is moderately bright (V=10.1), metal-poor ([Fe/H]$\approx-0.3$) G0V star with a much fainter M dwarf companion on a wide, eccentric orbit (=14.3 d). We have used near-infrared spectroscopy obtained with the SPIRou spectrograph to measure the semi-amplitude of the M dwarf's spectroscopic orbit, and high-precision photometry of the eclipse and transit from the CHEOPS and TESS space mis…
▽ More
EBLM J0113+31 is moderately bright (V=10.1), metal-poor ([Fe/H]$\approx-0.3$) G0V star with a much fainter M dwarf companion on a wide, eccentric orbit (=14.3 d). We have used near-infrared spectroscopy obtained with the SPIRou spectrograph to measure the semi-amplitude of the M dwarf's spectroscopic orbit, and high-precision photometry of the eclipse and transit from the CHEOPS and TESS space missions to measure the geometry of this binary system. From the combined analysis of these data together with previously published observations we obtain the following model-independent masses and radii: $M_1 = 1.029 \pm 0.025 M_{\odot}$, $M_2 = 0.197 \pm 0.003 M_{\odot}$, $R_1 = 1.417 \pm 0.014 R_{\odot}$, $R_2 = 0.215 \pm 0.002 R_{\odot}$. Using $R_1$ and the parallax from Gaia EDR3 we find that this star's angular diameter is $θ= 0.0745 \pm 0.0007$ mas. The apparent bolometric flux of the G0V star corrected for both extinction and the contribution from the M dwarf ($<0.2$ per cent) is ${\mathcal F}_{\oplus,0} = (2.62\pm 0.05)\times10^{-9}$ erg.cm$^{-2}$.s$^{-1}$. Hence, this G0V star has an effective temperature $T_{\rm eff,1} = 6124{\rm\,K} \pm 40{\rm \,K\,(rnd.)} \pm 10 {\rm \,K\,(sys.)}$. EBLM J0113+31 is an ideal benchmark star that can be used for "end-to-end" tests of the stellar parameters measured by large-scale spectroscopic surveys, or stellar parameters derived from asteroseismology with PLATO. The techniques developed here can be applied to many other eclipsing binaries in order to create a network of such benchmark stars.
△ Less
Submitted 6 May, 2022; v1 submitted 3 May, 2022;
originally announced May 2022.
