Why do Neural Decoders (GNNs) collapse into "syndrome-density shortcut learning" under correlated noise at high physical error rates?

Why do Neural Decoders (GNNs) collapse into "syndrome-density shortcut learning" under correlated noise at high physical error rates? Ask Question Asked 18 days ago Modified 11 days ago Viewed 49 times 0 Overview: I am building GNN-based decoders for the Surface Code (d=5) and benchmarking them against MWPM. To ensure the ML model doesn't lose hyperedge information via clique-expansion, I built an open-source pipeline (QEC Noise Factory) that orchestrates Stim/Sinter and extracts the DetectorErrorModel into an exact bipartite graph (preserving k>2 𝑘 > 2 hyperedges).When training under symmetric depolarizing noise, the GNN performs well. However, when introducing correlated crosstalk noise at higher physical error rates (e.g., p=0.04 𝑝 = 0.04 ), the GNN completely collapses into a phenomenon I call "K-leakage" or density shortcut learning. The Problem: Instead of learning the actual spatial topological correlations of the errors from the bipartite graph, the neural network acts as a trivial syndrome counter. It effectively computes the total number of active detectors (syndrome density, K 𝐾 ) and predicts based on the marginal probability P(logical_error|K) 𝑃 ( 𝑙 𝑜 𝑔 𝑖 𝑐 𝑎 𝑙 _ 𝑒 𝑟 𝑟 𝑜 𝑟 | 𝐾 ) , completely ignoring the graph structure.I confirmed this by creating a K-matched "scrambler" (shuffling detector spatial positions while keeping the total active count K 𝐾 constant). The model's AUROC on scrambled inputs perfectly matches its AUROC on clean inputs, yielding a "Topology Gain" of roughly zero. What I tried: I attempted to project out the K 𝐾 -collinear gradient components during backpropagation (using a custom Gradient Shield) to force the model to learn orthogonal topological features. While this worked for some random seeds (achieving organic clean representations), it caused a catastrophic topology collapse on other seeds. My Questions: Is this density-shortcut collapse a known fundamental barrier for purely structural neural decoders under high correlated noise (where SNR is low)? Are there established architectural patterns (e.g., contrastive Iso-K loss, or specific MPNN aggregations) that force a GNN to attend to the matching graph structure rather than the trivial density scalar near the threshold? Any insights, or pointers to literature dealing with this specific ML-QEC training pathology, would be highly appreciated. The reproducible pipeline and artifacts used to discover this are available in the linked repo: https://github.com/Moh-albataineh/qec-noise-factory.git error-correctionstimsurface-codemachine-learning Share Improve this question Follow edited Mar 5 at 19:54 Condo 2,2748 8 silver badges 31 31 bronze badges asked Feb 27 at 0:03 moe 1 Add a comment Know someone who can answer? Share a link to this question via email, Twitter, or Facebook. Your Answer Sign up or log in Sign up using Google Sign up using Email and Password Post as a guest Name Email Required, but never shown Post Your Answer By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy. Start asking to get answers Find the answer to your question by asking. Ask question Explore related questions error-correctionstimsurface-codemachine-learning See similar questions with these tags. The Overflow Blog Open source for awkward robots Domain expertise still wanted: the latest trends in AI-assisted knowledge for... Featured on Meta Logo updates to Stack Overflow's visual identity Hot Network Questions Relation between sign of electrode potential and oxidising/reducing nature of species Mirror stopped working Negative form of imperative nasai form Is there credible evidence that Pakistan’s ISI funds or supports the Khalistan movement abroad? Why is there a dash before "tu" in "Quel quartier préfères‑tu ?" What does it mean to be more electron dense and also slightly positively charged? The assumptions for causal inference seem obvious - why state them? For Pi Day, which expression is larger? Building an IR receiver circuit "Out of malice"? How should (0,-1) -- +([turn]45:2) be interpreted? What does "Probability content" of an interval mean? Protecting peonies from late freeze My basic budget tracker How to fish ethernet through ~1 ft run, instead of stapled coax? How can I diff two huge (250MB+) XML files with only minor changes? Is this spell balanced by the consequence? AI overview in normal Firefox window but not in Private Window Add two rational numbers... esoterically Hard position for Stockfish to analyze How do I equip previously unlocked costumes in SM64: Beyond the Cursed Mirror? proving that the extremal is independent of the coordinate system, can I swap differentials this way, since the derivative operator is linear? Which Elder Scrolls location is this? how to re-enable an ejected thumbdrive Question feed By continuing to use this website, you agree Stack Exchange can store cookies on your device and disclose information in accordance with our Cookie Policy. By exiting this window, default cookies will be accepted. To reject cookies, select an option from below. Customize settings Cookie Consent Preference Center When you visit any of our websites, it may store or retrieve information on your browser, mostly in the form of cookies. This information might be about you, your preferences, or your device and is mostly used to make the site work as you expect it to. The information does not usually directly identify you, but it can give you a more personalized experience. Because we respect your right to privacy, you can choose not to allow some types of cookies. Click on the different category headings to find out more and manage your preferences. Please note, blocking some types of cookies may impact your experience of the site and the services we are able to offer. Cookie Policy Accept all cookies Manage Consent Preferences Strictly Necessary Cookies Always Active These cookies are necessary for the website to function and cannot be switched off in our systems. They are usually only set in response to actions made by you which amount to a request for services, such as setting your privacy preferences, logging in or filling in forms. You can set your browser to block or alert you about these cookies, but some parts of the site will not then work. These cookies do not store any personally identifiable information. Targeting Cookies Targeting Cookies These cookies are used to make advertising messages more relevant to you and may be set through our site by us or by our advertising partners. They may be used to build a profile of your interests and show you relevant advertising on our site or on other sites. They do not store directly personal information, but are based on uniquely identifying your browser and internet device. Performance Cookies Performance Cookies These cookies allow us to count visits and traffic sources so we can measure and improve the performance of our site. They help us to know which pages are the most and least popular and see how visitors move around the site. All information these cookies collect is aggregated and therefore anonymous. If you do not allow these cookies we will not know when you have visited our site, and will not be able to monitor its performance. Functional Cookies Functional Cookies These cookies enable the website to provide enhanced functionality and personalisation. They may be set by us or by third party providers whose services we have added to our pages. If you do not allow these cookies then some or all of these services may not function properly. Cookie List Clear checkbox label label Apply Cancel Consent Leg.Interest checkbox label label checkbox label label checkbox label label Necessary cookies only Confirm My Choices