Interpreting deep learning models for weak lensing

Abstract

Deep Neural Networks (DNNs) are powerful algorithms that have been proven capable of extracting non-Gaussian information from weak lensing (WL) data sets. Understanding which features in the data determine the output of these nested, non-linear algorithms is an important but challenging task. We analyze a DNN that has been found in previous work to accurately recover cosmological parameters in simulated maps of the WL convergence ($\kappa$). We derive constraints on the cosmological parameter pair $({\Omega }_m,{\sigma }_8)$ from a combination of three commonly used WL statistics (power spectrum, lensing peaks, and Minkowski functionals), using ray-traced simulated $\kappa$ maps. We show that the network can improve the inferred parameter constraints relative to this combination by $20\%$ even in the presence of realistic levels of shape noise. We apply a series of well established saliency methods to interpret the DNN and find that the most relevant pixels are those with extreme $\kappa$ values. For noiseless maps, regions with negative $\kappa$ account for $86-69\%$ of the attribution of the DNN output, defined as the square of the saliency in input space. In the presence of shape nose, the attribution concentrates in high convergence regions, with $36-68\%$ of the attribution in regions with $\kappa >3{\sigma }_{\kappa }$.

arXiv comment12 pages, 7 figures, accepted for publication in PRD

References in corpus (39)

• SciPy 1.0--Fundamental Algorithms for Scientific Computing in Python
• Axiomatic Attribution for Deep Networks
• Weak Gravitational Lensing
• Wide-Field InfrarRed Survey Telescope-Astrophysics Focused Telescope Assets WFIRST-AFTA 2015 Report
• Why your model parameter confidences might be too optimistic -- unbiased estimation of the inverse covariance matrix
• Not Just a Black Box: Learning Important Features Through Propagating Activation Differences
• Cosmology from cosmic shear power spectra with Subaru Hyper Suprime-Cam first-year data
• Cosmology with cosmic shear observations: a review
• Visualizing Deep Neural Network Decisions: Prediction Difference Analysis
• Dark Energy Survey Year 1 Results: Constraints on Extended Cosmological Models from Galaxy Clustering and Weak Lensing
• Parameter inference with estimated covariance matrices
• The first-year shear catalog of the Subaru Hyper Suprime-Cam SSP Survey
• Cosmology with the shear-peak statistics
• Robust Morphological Measures for Large-Scale Structure in the Universe
• The (Un)reliability of saliency methods
• Cosmology Constraints from the Weak Lensing Peak Counts and the Power Spectrum in CFHTLenS
• Cosmology constraints from shear peak statistics in Dark Energy Survey Science Verification data
• Cosmological constraints with deep learning from KiDS-450 weak lensing maps
• KiDS-450: Cosmological Constraints from Weak Lensing Peak Statistics - II: Inference from Shear Peaks using N-body Simulations
• Probing Cosmology with Weak Lensing Peak Counts
• Massive Neutrinos Leave Fingerprints on Cosmic Voids
• Non-Gaussian information from weak lensing data via deep learning
• Emulating the CFHTLenS Weak Lensing data: Cosmological Constraints from moments and Minkowski functionals
• Cosmological Constraints From Weak Lensing Peak Statistics With CFHT Stripe 82 Survey
• Weak lensing cosmology with convolutional neural networks on noisy data
• Cosmological constraints from noisy convergence maps through deep learning
• Interpretation of Neural Networks is Fragile
• The Santiago-Harvard-Edinburgh-Durham void comparison II: unveiling the Vainshtein screening using weak lensing
• From Weak Lensing to non-Gaussianity via Minkowski Functionals
• Constraining neutrino mass with weak lensing Minkowski Functionals
• Weak Lensing Minima and Peaks: Cosmological Constraints and the Impact of Baryons
• On the dissection of degenerate cosmologies with machine learning
• Weak lensing by voids in weak lensing maps
• Baryon effects on void statistics in the EAGLE simulation
• Distinguishing general relativity and $f(R)$ gravity with the gravitational lensing Minkowski functionals
• Parameter inference and model comparison using theoretical predictions from noisy simulations
• Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps
• The matter fluctuation amplitude inferred from the weak lensing power spectrum and correlation function in CFHTLenS data
• Top-down Neural Attention by Excitation Backprop

Cited by in corpus (28)

• The CAMELS project: Cosmology and Astrophysics with MachinE Learning Simulations
• Large-scale dark matter simulations
• The DAWES review 10: The impact of deep learning for the analysis of galaxy surveys
• Cosmological Forecast for non-Gaussian Statistics in large-scale weak Lensing Surveys
• Learning cosmology and clustering with cosmic graphs
• Lossless, Scalable Implicit Likelihood Inference for Cosmological Fields
• Cosmology with one galaxy?
• The Cosmic Graph: Optimal Information Extraction from Large-Scale Structure using Catalogues
• Noise reduction for weak lensing mass mapping: An application of generative adversarial networks to Subaru Hyper Suprime-Cam first-year data
• Machine Learning for Observational Cosmology
• Simultaneously constraining cosmology and baryonic physics via deep learning from weak lensing
• Probability distribution function of the aperture mass field with large deviation theory
• Extracting cosmological parameters from N-body simulations using machine learning techniques
• Multifield Cosmology with Artificial Intelligence
• A comparative study of cosmological constraints from weak lensing using Convolutional Neural Networks
• Modeling assembly bias with machine learning and symbolic regression
• Comparing weak lensing peak counts in baryonic correction models to hydrodynamical simulations
• The Importance of Being Interpretable: Toward An Understandable Machine Learning Encoder for Galaxy Cluster Cosmology
• Cosmological constraints from low redshift 21 cm intensity mapping with machine learning
• Inpainting hydrodynamical maps with deep learning
• Accurate Kappa Reconstruction Algorithm for masked shear catalog (AKRA)
• Massive $\nu$s through the CNN lens: interpreting the field-level neutrino mass information in weak lensing
• AKRA 2.0: Accurate Kappa Reconstruction Algorithm for masked shear catalog
• Weak lensing peak statistics -- steepness versus height
• Differentiating Warm Dark Matter Models through 21cm Line Intensity Mapping: A Convolutional Neural Network Approach
• Interpreting Cosmological Information from Neural Networks in the Hydrodynamic Universe
• Assessing Galaxy Rotation Kinematics: Insights from Convolutional Neural Networks on Velocity Variations
• Interpretability of deep-learning methods applied to large-scale structure surveys