Neural Learning Methods TRIZ Cases
- Automated Proxy Task Design for Neural Architecture Search
Automated proxy task design tools minimize engineering effort by ranking candidates via correlation scores, reducing neural architecture search costs.
- Scalable Time Series Forecasting via Parallel Seasonal Encoders
A scalable deep neural network architecture uses parallel encoders to generate summary vectors for seasonal correlations.
- Variable Batch Sizing and Gradient Checkpoint Segments
Dynamic batch sizing adjusts training loads while gradient checkpointing discards intermediate activations to optimize memory usage.
- Neural Network Model Layer Segmentation for Storage Optimization
Segmenting neural network models into independent layers reduces storage space and bandwidth usage by transferring only modified components.
- Deep Learning Parameter Reuse via Validation
A deep learning parameter reuse method selects reusable weights through layer correspondence mapping and validation checks.
- Dynamic Tabu Tenure Adjustment in Combinatorial Optimization
A data processing device calculates an index value from evaluation function changes to determine suppression periods for state variables.
- Synthetic Data Generation for Visual Language Model Training
Text-to-image models generate synthetic training tuples, resolving limited dataset bias in visual language models.
- Multiscale Point Cloud Classification via Parallel Network Fusion
Parallel classification networks extract local and global features while a deep Q network fuses results to resolve automation versus accuracy contradictions.
- Interleaved Transformers With Latent Tokens
Interleaved transformers partition data tokens into groups and process them with adaptive latent tokens for global attention.
- Semi-Supervised Learning Model Training with Loss and Task Modules
A semi-supervised learning method trains a loss module to estimate label likelihoods and a task module using unlabelled data for prediction.
- Multi-layer RNN and CNN for Context Sensitive OCR
Nested RNN layers process CNN outputs to resolve recognition inaccuracies by leveraging hierarchical contextual information.
- ML Preflight System for Print Job Error Detection
Generative adversarial networks normalize print job data to identify subtle file errors before printing begins.
- Adversarial Noise Generator for Robust AI Feature Classification
An adversarial noise generator creates corrupted features to train a robust compressor that preserves useful data integrity.
- MNDR Model for Signal Quality Estimation in Optical Networks
A modular network model adapts to device conditions using shared encoders and specific decoders.
- GAN-Based Training Data Recovery for Pre-Trained Classifiers
A generative adversarial network generates simulated data samples to approximate original training inputs.
- Dialog-Based Music Recommendation System for Video Context
A dialog-based music recommendation system generates natural language suggestions tailored to video content using multi-modal analysis.
- Sparse DNN Weight Structure via Walsh-Hadamard Transform
Walsh-Hadamard transform enables sparse orthogonal kernels, reducing multiplications by 2.25 times while maintaining accuracy on IoT devices.
- Multi-Modal Language Model Reconstruction Feedback Loop
A multi-modal language model employs a decoder network to reconstruct media tokens, reducing hallucinations and contradictions in text responses.
- Generative AI Message Suggestion System with Scoring Model
A generative message suggestion system uses a scoring model to rank content based on user profiles and communication history.
- Masked Multi-Step Forecasting for Time Series Accuracy
A masked multi-step forecasting system applies neural network masks to historical sequences for direct prediction.
- Conversation Content Modification via ML Emotional Alignment
Machine learning models process biometric and video data to adjust text tone, resolving the trade-off between detection accuracy and processing time.
- Personalized Federated Learning via Partial Parameter Segmentation
Segmenting neural network parameters into shared global and private local sets reduces communication overhead while maintaining personalized model accuracy.
- Virtual Batches for Neural Network Signal Preservation
Segmenting variable-length sequences into fixed-size windows enables virtual batch construction, preserving temporal signal integrity during RNN model training.
- Vector Neural Network Feature Spectrum Evaluation
A vector neural network model extracts feature spectra from intermediate layers to evaluate target data using general-purpose training sets.
- Neural Network Feature Detectors Overfitting Control
Random switch disables feature detectors during neural network training, preventing complex co-adaptations and improving generalization performance.
- Embedding Model Training via Local-Global View Contrast
Contrasting local and global view outputs from dual embedding models reduces paired dataset costs while maintaining multi-modal retrieval accuracy.
- Contextual Hypernetwork for Predicting New Model Parameters
A contextual hypernetwork predicts new parameters for deep learning models without additional training.
- Graph Neural Network Warm-Start for Newton-Raphson Power Flow
A graph neural network generates a warm-start solution to accelerate the Newton-Raphson algorithm.
- Radar DoA Estimation via ML Covariance Matrix
A radar system uses a machine learning model to generate additional spatial covariance matrices for direction of arrival estimation.
- Automated Training Data Generation for Deep Learning Models
A system generates training data by detecting original data types and applying specific processing techniques.
- Bi-directional Neural Network for Unpaired Image Translation
A bi-directional neural network translates images between domains using harmonic and circularity loss functions to enforce consistency.