[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-327427-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-327427-en":130},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","dynamic-neural-graph-encoding-of-inference-processes-in-deep-weight-space","Dynamic Neural Graph Encoding of Inference Processes in Deep Weight Space","","Neural networks can be used as implicit data representations, motivating machine-learning methods that analyze other networks’ weight spaces. Efficiently handling high-dimensional weight spaces remains difficult, especially when inference proceeds sequentially layer by layer. This work proposes dynamic neural graphs and a Dynamic Neural Graph Encoder (DNG-Encoder) to model temporal evolution across the forward pass. It further builds INR2JLS to learn a joint latent space, improving INR classification by about 10% on CIFAR-100-INR.",{"@graph":14,"@context":73},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/dynamic-neural-graph-encoding-of-inference-processes-in-deep-weight-space/327427/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/dynamic-neural-graph-encoding-of-inference-processes-in-deep-weight-space/327427.png","ImageObject",300,407,{"name":42,"@type":43},"Ivy","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-10-29","2026-09-21",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",9,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What problem does the work address in neural network weight-space analysis?","Question",{"text":63,"@type":64},"It addresses the challenge of efficiently processing high-dimensional neural weight spaces while accounting for the sequential, layer-by-layer nature of inference.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"How does DNG-Encoder differ from prior graph-based weight-space models?",{"text":68,"@type":64},"It uses dynamic graphs to capture temporal dynamics during the forward pass, whereas prior approaches mostly rely on static graphs that process the entire graph in one pass.",{"name":70,"@type":61,"acceptedAnswer":71},"What is INR2JLS and what downstream benefit does it target?",{"text":72,"@type":64},"INR2JLS uses DNG-Encoder to learn a joint latent space between deep weights and the original data, improving downstream applications such as classifying implicit neural representations.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},327427,1790633216,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,111,115,119,122,126],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":107,"doc_module":4,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},7,"Healthcare",40,"healthcare",{"id":112,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":113,"slug":114},8,30,"research-report",{"id":55,"doc_module":4,"doc_module_name":25,"category_name":116,"show_sort_weight":117,"slug":118},"Religion & Spirituality",20,"religion-spirituality",{"id":117,"doc_module":4,"doc_module_name":25,"category_name":120,"show_sort_weight":117,"slug":121},"World Cup","world-cup",{"id":123,"doc_module":4,"doc_module_name":25,"category_name":124,"show_sort_weight":123,"slug":125},10,"Lifestyle","lifestyle",{"id":127,"doc_module":4,"doc_module_name":25,"category_name":128,"show_sort_weight":97,"slug":129},19,"General","general",{"code":4,"msg":82,"data":131},{"doc_id":79,"user_id":132,"nickname":42,"user_avatar":133,"doc_module":4,"category_id":112,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":139,"language":140,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":12,"update_tm":144,"read_time":145,"preview_page_count":4},549758252649,"https://ap-avatar.wpscdn.com/avatar/8000253669c5317157?_k=1778319167496531819","arXiv :2607 .02 166v 1 [ cs .LG] 2 Jul 2026  \nDynamic Neural Graph Encoding of Inference Processes in Deep Weight Space  \nDi Wu  \nUniversity of Toronto  \nNational University of Singapore  \nHuan Liu∗  \nMcMaster University  \n[E1538518@u. nus. edu](E1538518@u. nus. edu)  \n[liuh127@outlook. com](liuh127@outlook. com)  \nZhixiang Chi  \nUniversity of Toronto  \nYuanhao Yu  \nMcMaster University  \nKonstantinos N. Plataniotis  \nUniversity of Toronto  \nYang Wang  \nConcordia University  \nReviewed on OpenReview: [https: // openreview. net/ forum? id= 4fweEyVYLF](https: // openreview. net/ forum? id= 4fweEyVYLF)  \nAbstract  \nThe rapid advancements in using neural networks as implicit data representations have attracted significant interest in developing machine learning methods that analyze and process the weight spaces of other neural networks. However, efficiently handling these highdimensional weight spaces remains challenging. Existing methods often overlook the sequential nature of layer-by-layer processing in neural network inference. In this work, we propose a novel approach using dynamic graphs to represent neural network parameters, capturing the temporal dynamics of inference. Our Dynamic Neural Graph Encoder (DNG-Encoder) processes these graphs, preserving the sequential nature of neural processing. Additionally, we also leverage DNG-Encoder to develop INR2JLS (Implicit Neural Representation to Joint Latent Space) for facilitate downstream applications, such as classifying Implicit Neural Representations (INRs) . Our approach demonstrates significant improvements across multiple tasks, surpassing the state-of-the-art INR classification accuracy by approximately 10% on the CIFAR-100-INR. Our code is available at [https://github.com/dddiowww/DNG](https://github.com/dddiowww/DNG).  \n1 Introduction  \nDeep neural networks have demonstrated superb capability in addressing real-world problems in fields such as computer vision, natural language processing, and the natural sciences. While it is generally used for learning patterns from data, recent studies have expanded its scope by treating neural networks themselves as inputs, enabling tasks such as optimizing networks (Metz et al., 2022), predicting the labels of the data encoded in implicit neural representations (Dupont et al., 2022), and generating or modifying their weights to alter functionality (Schürholt et al., 2022) . However, processing these weight spaces presents considerable challenges due to their complex, high-dimensional nature.  \n∗ Research Lead. Corresponding Author.  \nTo address the difficulty, some existing methods propose to narrow the effective weight space using a restricted training process (Bauer et al., 2023; De Luigi et al., 2023) . However, this neglects the crucial permutation symmetry property of the neural network weights, i. e . , neurons within a layer can be rearranged without altering the network’s function (Hecht-Nielsen, 1989) . Overlooking the permutation symmetry can significantly increase the search space for optimal parameters of processing network, resulting in reduced generalization and unsatisfactory performance. By observing this, recent works (Navon et al., 2023; Zhou et al., 2024a;b) build permutation equivariant weight-space models named neural functionals. Unfortunately, these methods need manual adaptation for each new architecture, and a single model can only handle one fixed architecture. To enable processing heterogeneous architectures, Kofinas et al. (2024) and Lim et al. (2024) introduce to model neural network weights as graphs, which links neural network parameters similarly to a computation graph. These methods, while innovative, predominantly employ static graphs. This static representation allows GNNs to process the entire graph in a single pass. However, such an approach overlooks a critical aspect of neural network behavior during inference: the sequential nature of layer-by-layer processing. Neural networks, by design, perfo","cbCaivyDKIgakmW7","https://ap.wps.com/l/cbCaivyDKIgakmW7","pdf",2470948,28,"English","# Abstract\n# 1 Introduction\n# 2 Related Work\n## Dynamic Graphs","[{\"question\":\"What problem does the work address in neural network weight-space analysis?\",\"answer\":\"It addresses the challenge of efficiently processing high-dimensional neural weight spaces while accounting for the sequential, layer-by-layer nature of inference.\"},{\"question\":\"How does DNG-Encoder differ from prior graph-based weight-space models?\",\"answer\":\"It uses dynamic graphs to capture temporal dynamics during the forward pass, whereas prior approaches mostly rely on static graphs that process the entire graph in one pass.\"},{\"question\":\"What is INR2JLS and what downstream benefit does it target?\",\"answer\":\"INR2JLS uses DNG-Encoder to learn a joint latent space between deep weights and the original data, improving downstream applications such as classifying implicit neural representations.\"}]","Dynamic Neural Graph Encoding of Inference Processes in Deep Weight Space | PDF",1789977180,71]