Neural Architectures TRIZ Cases
- LSTM Neural Network Grapheme to Phoneme Conversion
An LSTM neural network generates initial phoneme sequences from grapheme inputs for accurate prediction.
- Neural Network Architecture Reduction via Weight Transformation
Reduced space transformation eliminates unimportant neurons and minimizes repetitive training cycles while maintaining high prediction accuracy.
- Temporal Prompt Generator for Neural Network Domain Adaptation
A temporal prompt generator adapts frozen backbone neural networks to new domains by producing dynamic prompts that capture underlying data dynamics.
- Binary Tensor Scale Factor Pairs for Neural Network Compression
Binary tensor and scale factor pairs approximate trained weights to reduce computational burden in neural networks.
- Three-Terminal Variable Resistance Element for Spiking Neural Networks
A three-terminal variable resistance element enables spiking neural network implementation through independent control of resistance values.
- View-Class Operator Subgraph Optimization for Tensor Data Contiguity
Optimizes tensor data access by constructing a view-class operator subgraph that reduces memory moving and operator calling overhead.