Get Started
Home
Topics
Search
Library
Research questionHow can we characterize when deep ReLU networks with different architectures or parameters compute the same function on the unit cube?Different deep ReLU architectures and parameter settings can induce exactly the same input-output map, making network identification non-unique. The difficulty is to characterize all such equivalent representations, including those differing in their layer structure.
AI
Machine Learning
Neural and Evolutionary Computing
Research Paper
Technology
Latest papersRecent research connected to this question, newest first.Complete Identification of Deep ReLU Networks through Łukasiewicz LogicApplies to non-degenerate deep ReLU networks whose functions are considered on the unit cube. The source covers integer, rational, and real weights and biases and supplies algorithms for extracting and reconstructing layered symbolic representations, but the stated completeness result is limited to this setting.research paper · Sep 3, 2026
Related questions
How do input dimension and network depth affect the tractability of verifying properties of ReLU networks?How does the leaky-ReLU parameter alter critical-point bifurcations and symmetry-breaking stability in shallow networks?Does reusing Transformer layers improve language-model quality when parameter, compute, and KV-cache budgets are matched?How do initialization correlation decay and feature-function Hermite rank determine residual networks’ scaling and continuous-depth limit?