Neural Learning Methods TRIZ Cases
- Early-Exit Machine Learning Model Training Method
Training untrained classifiers first eliminates adverse gradient effects, improving inference accuracy while maintaining speed.
- Neural Network 3D Image Classification via 2D Slice Grouping
A neural network classifies three-dimensional images by processing two-dimensional slices in grouped sequences matching standard color image inputs.
- Multi-Input Speech Synthesis Model for Chinese Pronunciation Accuracy
A speech synthesis model fuses syllable, phoneme, and character inputs to improve pronunciation accuracy.
- Neural Network Waveform Signal Processing for Noise Reduction
A neural network generates time-series data to separate noise from elastic wave signals using a learner that updates parameters based on a loss function.
- Fixed Recurrent Matrices for Neural Network Dimensionality
Segmenting weight matrices into fixed and learned parts reduces computational complexity while maintaining high-dimensional transformations.
- Automatic Pipeline Parameter Tuning via Iterative Feedback
An automatic pipeline tuning system uses iterative feedback to optimize parameters without manual intervention.
- Few-Shot Learning Method Using Multi-Task Neural Network
A few-shot learning method uses a multi-task neural network to improve feature representation.
- Target Detection Using Dilated Convolution for Robustness
Dilated convolution expands receptive fields in image features to improve detection robustness while reducing processing time.
- Autoencoder Anomaly Detection via Genetic Algorithm Optimization
Genetic algorithm optimization trains non-mirrored autoencoders to detect anomalies with reduced computing resource consumption.
- Dual-Decoder ML Model for Context-Specific Question Diversity
Dual decoders process target questions with context to generate diverse product queries, resolving generic output issues in digital marketplaces.
- One-Dimensional Convolution Neural Network for Real-Time Signal Classification
A signal-analyzing neural network uses one-dimensional convolutional layers to classify parameter-varying electromagnetic signals.
- Structured Data Normalization for Generative AI Model Training
Computing platform normalizes structured historical information to train generative AI models using foundational model features.
- Neural Network Heuristics for Mixed Integer Program Solving
A deep neural network generates partial assignments to accelerate mixed integer program solving.
- Neural Architecture Search Using Gradient-Based Sub-Space Optimization
A neural architecture search method uses gradient-based sub-space optimization to identify candidate layers for efficient network design.
- Neural Network Weight Tensor Compression via Unified Tensor Basis
A unified tensor basis parametrizes weight tensors to reduce storage coefficients.