{"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/analysing-neural-network-topologies-a-game","title":"Analysing Neural Network Topologies: a Game Theoretic Approach","arxiv_id":"1904.08166","date":"2019-04-17","proceeding":null,"authors":["Julian Stier","Gabriele Gianini","Michael Granitzer","Konstantin Ziegler"],"abstract":"Artificial Neural Networks have shown impressive success in very different\napplication cases. Choosing a proper network architecture is a critical\ndecision for a network's success, usually done in a manual manner. As a\nstraightforward strategy, large, mostly fully connected architectures are\nselected, thereby relying on a good optimization strategy to find proper\nweights while at the same time avoiding overfitting. However, large parts of\nthe final network are redundant. In the best case, large parts of the network\nbecome simply irrelevant for later inferencing. In the worst case, highly\nparameterized architectures hinder proper optimization and allow the easy\ncreation of adverserial examples fooling the network. A first step in removing\nirrelevant architectural parts lies in identifying those parts, which requires\nmeasuring the contribution of individual components such as neurons. In\nprevious work, heuristics based on using the weight distribution of a neuron as\ncontribution measure have shown some success, but do not provide a proper\ntheoretical understanding. Therefore, in our work we investigate game theoretic\nmeasures, namely the Shapley value (SV), in order to separate relevant from\nirrelevant parts of an artificial neural network. We begin by designing a\ncoalitional game for an artificial neural network, where neurons form\ncoalitions and the average contributions of neurons to coalitions yield to the\nShapley value. In order to measure how well the Shapley value measures the\ncontribution of individual neurons, we remove low-contributing neurons and\nmeasure its impact on the network performance. In our experiments we show that\nthe Shapley value outperforms other heuristics for measuring the contribution\nof neurons.","url_abs":"http://arxiv.org/abs/1904.08166v1","url_pdf":"http://arxiv.org/pdf/1904.08166v1.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":[],"tasks":[{"task_slug":"network-pruning","task_name":"Network Pruning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/network-pruning-on-mnist","task":"Network Pruning","dataset":"MNIST","model":"FFN-ShapleyPruned","rank_in_archive_order":1,"of":1,"metrics":{"Avg #Steps":"12.05"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1904.08166","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}