Research map / Architectures and efficiency / RNNs and normalization
Real-Valued Backpropagation is Unsuitable for Complex-Valued Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase hidden width of complex MLP (CReLU, CSigmoid, Ctanh) from 2^5 to 2^13 | complex-valued MLP under real-valued backprop · synthetic inputs, NTK at initialization | relative Frobenius norm to real NTK | Helped |
| increase hidden width from 100 to 1000 (CReLU, modified modReLU) | complex-valued MLP · synthetic z=1-i vs z' on unit circle | concentration of empirical complex NTK around real NTK | Helped |
| use zReLU activation that violates the theorem's condition instead of CReLU/modified modReLU (ablation) | complex-valued MLP · synthetic inputs, NTK at initialization | gap between complex and real NTK | Hurt |
| modified modReLU with different threshold rho (0.5 vs 0.3) (ablation) | complex-valued MLP · synthetic inputs, width 2^5 to 2^13 | relative Frobenius norm to real NTK | No clear effect |
| increase width from 2^5 to 2^10 during training (t=1000) on MNIST subset, CReLU | complex-valued MLP · MNIST subset of 128 samples | relative Frobenius norm to real NTK | Helped |
| go from initialization (t=0) to training (t=1000) (ablation) | complex-valued MLP vs real NTK · MNIST subset, L=1 and L=2 | relative Frobenius norm to real NTK | No clear effect |
| one vs two hidden layers (ablation) | complex-valued MLP · MNIST subset and synthetic | convergence to real NTK | No clear effect |
Closest papers by meaning
- Complex-valued Neurons Can Learn More but Slower than Real-valued Neurons via Gradient Descent (NeurIPS 2023 · similarity 0.86)
- Better NTK Conditioning: A Free Lunch from (ReLU) Nonlinear Activation in Wide Neural Networks (NeurIPS 2025 · similarity 0.85)
- The Surprising Effectiveness of Infinite-Width NTKs for Characterizing and Improving Model Training (AAAI 2025 · similarity 0.84)
- Feature Learning Dynamics in Infinite-Depth Neural Networks (NeurIPS 2026 · similarity 0.84)
- Understanding the Evolution of the Neural Tangent Kernel at the Edge of Stability (NeurIPS 2025 · similarity 0.84)
- On the Impacts of the Random Initialization in the Neural Tangent Kernel Theory (NeurIPS 2024 · 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.