Neural Learning Methods TRIZ Cases
- Generative Model Output Length Compliance With Synthetic Training
Generative models are evaluated against word, sentence, and paragraph limits, then retrained with synthetic examples that satisfy those constraints.
- Information Processing With Define-by-Run Graph Generation
Define-by-Run execution builds backward-reference structures from forward code, supporting dynamic graphs, memory efficiency, and complex neural networks.
- Multi-Task Model Training with Shared and Task-Specific Sub-Models
Shared and task-specific sub-models train across multiple datasets, reducing the cost of separate neural-network training.
- Visually Aware Layout Generation With Multi-Domain Diffusion
Conventional layout generators overlook image appearance; multi-domain diffusion uses visual features to improve saliency and diversity.
- 1.5-Branch Bipolar Morphological Networks With Layer-by-Layer Distillation
Layer-by-layer distillation converts classical networks into 1.5-branch bipolar models while retaining efficient hardware inference.
- Difference-Signal Classifier Training for Automated Label Generation
Difference signals from generated and input signals automate classifier labels, reducing manual annotation time while supporting larger training and testing datasets.
- Image File Section Analysis with Deep Learning for Malware Detection
Neural networks assess selected image-file sections to expose disguised malware while limiting false positives and processing time.
- Machine Learning Model Training With Staged Bias Mitigation
Pre-processing, in-processing, and post-processing checks detect bias and apply targeted corrections while conserving computational resources.
- Edge DNN Structured Pruning at Initialization for Rapid Deployment
Unstructured pruning can preserve accuracy without improving runtime; SPaI uses layer sensitivity to build smaller, faster edge DNNs.
- Concurrent Phoneme Processing for Faster RNN Speech Synthesis
An RNN concurrently processes adjacent phoneme features, reducing iterative TTS processing time and speeding audio synthesis.
- Thermally Adaptive CT Scan Sequencing Without Fixed Delays
Temperature-aware scan ordering helps CT scanners limit thermal stress without relying on excessive delays between protocols.
- Blockchain-Enabled Federated Learning for IoT Intrusion Detection
Local IoT clients train on distributed data while blockchain-secured model sharing improves intrusion detection without exposing personal information.
- Neural Network Average Quantization for Even Weight Distribution
Uneven weights can reduce neural-network accuracy after quantization; Gini-regularized training evens distributions before integer conversion.
- Watchlist Identity Resolution Using Graph-Based Machine Learning
Graph-based clustering and semantic modeling resolve aliases and PII errors, reducing false positives and negatives in watchlist matching.
- Ordinal Mapping Neural Networks With Centroid-Free Loss
A centroid-free ordinal mapping loss measures class distances and overlap, enabling ordered classification without hyperplane centroid calculations.
- Channel State Feedback Neural Networks with Classifier Weight Selection
A UE selects pretrained neural-network weights with a classifier as channel conditions change, reducing re-training latency for feedback compression.
- Infinite-Width Neural Network Hyperparameter Transfer for Low-Cost Tuning
Tuning a narrow neural network and transferring its hyperparameters reduces the computation and energy needed to train billion-parameter models.