Neural Learning Methods TRIZ Cases
- GAN Speech Synthesis Model Reducing Acoustic Losses
Joint generator and discriminator training reduces acoustic losses in synthesized audio, improving voice quality beyond traditional separate training methods.
- Convolutional Neural Network Feature Map Decomposition for Accuracy and Overhead
A convolutional neural network decomposes feature maps into frequency subsets to extract features via learned upsampling.
- Multi-Party Communication Response Timing Using LLM Segmentation
An instruct-based language model detects response indicators in multi-party conversations to trigger a generative model for support replies.
- Unsupervised Pattern Discovery via Dynamic Graph Embeddings
Clustering algorithms separate noise from signal in dynamic graph embeddings, resolving the trade-off between predictive capability and interpretability.
- Hierarchical Agent System for Modular Gesture Control
Segmenting gesture control into independent agents reduces model complexity while enabling modular training across computing devices.
- Deep Neural Network Activation Function Selection
System selects activation functions based on industry type to train multiple hidden layers without manual configuration.
- Federated Learning Feature Normalization Layer
Normalization layers stabilize feature vectors across clients, resolving data heterogeneity issues in federated learning systems.
- Anomaly Detection in Unknown Domains Using Compressed Data
Neural network generates content-irrelevant and domain-irrelevant latent codes to reconstruct input data for anomaly detection.
- AI Model Information Erasure via Retraining and Tensorization
Tensorization compresses the model while retraining preserves performance, resolving privacy trade-offs.
- Multitouch Sensor Noise Separation Using GRU and CNN
A multitouch sensor system uses a gated recurrent unit and convolutional neural network to process touch data signals.
- Multi-View Image Processing System for Medical Imaging
A multi-view image processing system generates replacement images using AI models trained on time-adjacent sets to fill missing data gaps.
- AI Model Evaluation Metrics for Imbalanced Data
Cross F1 and Macro F1 scores resolve misleading accuracy in imbalanced data by weighting minority class performance.
- Neural Network Representation Learning for Complex Data Clustering
A neural network converts input data into representation vectors to enable effective clustering of complex datasets.
- Neural Network Incremental Learning via Probabilistic Training Set Selection
Probabilistic training set selection weights examples by recency to build efficient incremental learning subsets for neural networks.
- Federated Learning for Physical Layer Signal Testing
Segmented federated learning reduces device complexity while improving network stability and measurement accuracy.
- Expert-Choice Routing in Mixture-of-Experts Neural Networks
Expert subnetworks select input elements to resolve load imbalance and improve computational efficiency during neural network training.
- Training Data Generation Device for Feature Imbalance
A training data generation device classifies input samples into similarity groups to convert majority data while preserving minority feature proportions.
- Integrated Teacher-Student Neural Network for Efficient Inference
An integrated teacher-student neural network uses intermediate feature maps to generate inferences across distributed platforms.
- Image Recognition Model Training via Confidence-Based Sample Selection
A method selects untagged image samples within a preset confidence interval for manual labeling to update the training dataset.
- Wireless Motion Detection via Neural Network Signal Analysis
Neural network system analyzes wireless signal properties to detect motion, reducing false-positive rates by distinguishing human movement from other activity.
- Causal Discovery and Missing Value Imputation via Graph Neural Networks
Unified framework combines missing value imputation with causal discovery by tuning edge probabilities in a graph neural network.
- Neural Network Critical Neuron Selection via Anomaly Detection
An anomaly detection system identifies critical neurons during early training stages for strategic pruning.
- Autoencoder Latent Variable Extraction via Mutual Information Minimization
A machine learning model extracts interpretable latent variables from high-dimensional data using a unified cost function.
- 3D Object Detection Using Pseudo-Labels
Student neural networks trained with pseudo-labels resolve the contradiction between detection accuracy and data acquisition time.
- Synthetic Data Training for Robotic Depalletizing
Virtual image generation automates labeling and trains neural networks without manual data collection, reducing preparation time and cost.
- Complex Waveform Correction via Real Imaginary Partitioning
Separating complex waveforms into real and imaginary components enables independent neural analysis, reducing computing time while maintaining high accuracy.
- Negative Feedback Network for Fast Target Propagation
A negative feedback network perturbs output signals to rapidly determine local target outputs for weight updates.
- Supervised VAE Latent Space Optimization for Data Generation
A supervised variational autoencoder learns input distributions and target relationships to generate realistic optimal samples.
- Autoencoder-Based Weight Initialization for Neural Networks
Data-driven weight initialization uses autoencoder loss to determine neural network parameters, avoiding vanishing or exploding gradients during training.
- Neural Network Model Compression via Block Segmentation and Parameter Tuning
A neural network compression method segments trained models into blocks to enable user-configurable parameter tuning for efficient deployment.
- Task-Agnostic Bipartite Graph Node Representations
Generating task-agnostic node representations via augmented neighborhoods in bipartite graphs.
- Time-Series-Language Model Tokenization for Domain Adaptation
Encoder-decoder neural network tokenizes time-series data into discrete embeddings for language model processing.