{"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/hierarchical-interpretations-for-neural","title":"Hierarchical interpretations for neural network predictions","arxiv_id":"1806.05337","date":"2018-06-14","proceeding":"ICLR 2019 5","authors":["Chandan Singh","W. James Murdoch","Bin Yu"],"abstract":"Deep neural networks (DNNs) have achieved impressive predictive performance\ndue to their ability to learn complex, non-linear relationships between\nvariables. However, the inability to effectively visualize these relationships\nhas led to DNNs being characterized as black boxes and consequently limited\ntheir applications. To ameliorate this problem, we introduce the use of\nhierarchical interpretations to explain DNN predictions through our proposed\nmethod, agglomerative contextual decomposition (ACD). Given a prediction from a\ntrained DNN, ACD produces a hierarchical clustering of the input features,\nalong with the contribution of each cluster to the final prediction. This\nhierarchy is optimized to identify clusters of features that the DNN learned\nare predictive. Using examples from Stanford Sentiment Treebank and ImageNet,\nwe show that ACD is effective at diagnosing incorrect predictions and\nidentifying dataset bias. Through human experiments, we demonstrate that ACD\nenables users both to identify the more accurate of two DNNs and to better\ntrust a DNN's outputs. We also find that ACD's hierarchy is largely robust to\nadversarial perturbations, implying that it captures fundamental aspects of the\ninput and ignores spurious noise.","url_abs":"http://arxiv.org/abs/1806.05337v2","url_pdf":"http://arxiv.org/pdf/1806.05337v2.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":"hierarchical-interpretations-for-neural","repo_url":"https://github.com/csinva/hierarchical-dnn-interpretations","is_official":1,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"clustering","task_name":"Clustering"},{"task_slug":"feature-importance","task_name":"Feature Importance"},{"task_slug":"interpretable-machine-learning","task_name":"Interpretable Machine Learning"}],"methods":[{"method_slug":"hierarchical-dnn-interpretations","method_name":"Agglomerative Contextual Decomposition"}],"datasets_introduced":[],"methods_introduced":[{"slug":"hierarchical-dnn-interpretations","name":"Agglomerative Contextual Decomposition","full_name":"Agglomerative Contextual Decomposition"}],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1806.05337","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}