Neural Learning Methods TRIZ Cases
- Neural Network Rotor Alloy Design System
A neural network correlates microstructural features with material properties to design optimized rotor alloys.
- Vector Circuit Scalar Operations Accelerator
Accelerator circuit executes vector operations via scalar instructions, reducing CPU bandwidth consumption.
- Neural Network Training via Output Variable Exclusion
A neural network training method excludes a subset of output variables from backpropagation to focus learning on broader features.
- Drift Pattern Calibration for AI Prediction Models
A drift pattern creating unit detects latent factors in learning data to generate specific drift patterns for model calibration.
- Neural Network Pre-training for User Engagement Prediction
Pre-training on unlabeled data improves prediction accuracy while reducing training time.
- Low Rank Neural Network Training via SVD Initialization
A two-phase training technique applies singular value decomposition initialization to factorize neural network weight matrices into low-rank structures.
- Distributed Subclassification Model for Edge Terminals
Segmenting large models into specialized subclassification units distributes computational load and preserves accuracy despite terminal failures.
- Neural Architecture Search with Dynamic Time Budgets
Dynamic time budgets replace fixed computational steps in neural architecture search, enabling faster training convergence and improved generalization.
- Rule Encoder Integrating Symbolic Knowledge Into Deep Learning
A rule encoder corrects embedded representations using symbolic information to generate predictions.