Transformation Importance with Applications to Cosmology

• DOI: 10.48550/arxiv.2003.01926
• Publication Date: 2020-01-01
• Authors: Chandan Singh; Woo-Seok Ha; François Lanusse; Vanessa Boehm; Jia Liu; Bin Yu
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; FOS: Physical sciences; Machine Learning (stat.ML); 02 engineering and technology; [STAT.OT]Statistics [stat]/Other Statistics [stat.ML]; Machine Learning (cs.LG); [INFO.INFO-LG]Computer Science [cs]/Machine Learning [cs.LG]; Statistics - Machine Learning; 0202 electrical engineering, electronic engineering, information engineering; Astrophysics - Instrumentation and Methods for Astrophysics; Instrumentation and Methods for Astrophysics (astro-ph.IM)

Machine learning lies at the heart of new possibilities for scientific discovery, knowledge generation, and artificial intelligence. Its potential benefits to these fields requires going beyond predictive accuracy focusing on interpretability. In particular, many problems require interpretations in a domain-specific interpretable feature space (e.g. frequency domain) whereas attributions raw features pixel space) may be unintelligible or even misleading. To address this challenge, we propose TRIM (TRansformation IMportance), novel approach which attributes importances transformed can applied post-hoc fully trained model. is motivated by cosmological parameter estimation problem using deep neural networks (DNNs) simulated data, but it generally applicable across domains/models combined with any local interpretation method. our cosmology example, combining contextual decomposition shows promising results identifying frequencies DNN uses, helping cosmologists understand validate that model learns appropriate physical rather than simulation artifacts.