Researchers have developed CNet, a C++/CUDA framework designed for training deep complex-valued neural networks (CVNNs) and optimizing complex-valued functions using Wirtinger derivatives. This framework adopts a physics-native approach, treating neural networks as cascades of complex operations and classification as a Born-rule measurement. CNet includes CPU references and CUDA kernels for its layers, with a computation graph cloned across batches for GPU execution. The framework also incorporates signal-processing primitives for learnable complex convolutional networks and a true-Adam optimizer with a reduced-memory inference mode. Initial studies show that a complex-valued, FNet-style causal sequence model built with CNet matches or surpasses its real-valued counterpart in character-level language modeling, achieving converged quality in less than half the training time. AI
IMPACT This framework could enable new approaches to modeling complex data in fields like signal processing and physics.
RANK_REASON The cluster describes a new framework and associated research paper for complex-valued deep learning. [lever_c_demoted from research: ic=1 ai=1.0]
- Adam
- Auguste Hadamard
- Bluestein
- CNET
- CUDA
- density functional theory
- FNET
- RML2016.10a
- Wirtinger derivative
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
