{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/computationally-efficient-measures-of","title":"Computationally Efficient Measures of Internal Neuron Importance","arxiv_id":"1807.09946","date":"2018-07-26","proceeding":null,"authors":["Avanti Shrikumar","Jocelin Su","Anshul Kundaje"],"abstract":"The challenge of assigning importance to individual neurons in a network is\nof interest when interpreting deep learning models. In recent work, Dhamdhere\net al. proposed Total Conductance, a \"natural refinement of Integrated\nGradients\" for attributing importance to internal neurons. Unfortunately, the\nauthors found that calculating conductance in tensorflow required the addition\nof several custom gradient operators and did not scale well. In this work, we\nshow that the formula for Total Conductance is mathematically equivalent to\nPath Integrated Gradients computed on a hidden layer in the network. We provide\na scalable implementation of Total Conductance using standard tensorflow\ngradient operators that we call Neuron Integrated Gradients. We compare Neuron\nIntegrated Gradients to DeepLIFT, a pre-existing computationally efficient\napproach that is applicable to calculating internal neuron importance. We find\nthat DeepLIFT produces strong empirical results and is faster to compute, but\nbecause it lacks the theoretical properties of Neuron Integrated Gradients, it\nmay not always be preferred in practice. Colab notebook reproducing results:\nhttp://bit.ly/neuronintegratedgradients","url_abs":"http://arxiv.org/abs/1807.09946v1","url_pdf":"http://arxiv.org/pdf/1807.09946v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"computationally-efficient-measures-of","repo_url":"https://github.com/pytorch/captum","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"BSD-3-Clause"}}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1807.09946","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}