Research map / Architectures and efficiency / RNNs and normalization
Neural Spectral Capacity: An Architectural Quantity from Network Specification Alone
CHENYU ZHU, Ruoyu Zhao, Zhichao Lu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Neural Spectral Capacity (NSC), a closed-form architectural quantity computed from network specification alone, as a structure-aware alternative to #Params and #FLOPs for Transformer design and compression decisions.
Problem
Standard scalars #Params and #FLOPs measure size and compute but ignore architectural structure (depth-width, head, and FFN allocation), so architectures with identical budgets receive identical scores despite behaving differently.
Approach
A closed-form spectral capacity measure computed from the network specification alone (no training), intended to distinguish architectures with equal parameter budgets but different structural allocations.
Experimental setup
Baselines: #Params #FLOPs
Closest papers by meaning
- Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology (NeurIPS 2026 · similarity 0.86)
- Principled Architecture-aware Scaling of Hyperparameters (ICLR 2024 · similarity 0.85)
- Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers (ICML 2025 · similarity 0.85)
- On the Expressive Power of Spectral Invariant Graph Neural Networks (ICML 2024 · similarity 0.84)
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory (NeurIPS 2026 · similarity 0.84)
- Memory Layers at Scale (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
