Neural Learning Methods TRIZ Cases
- Neural Network Augmentation for Sequence Inference Stability
Augments neural networks by combining layer outputs to infer outcomes from image sequences without retraining.
- Contrastive Embedding for Bayesian Optimization
A contrastive learning approach generates alternative embeddings by creating structurally similar sample pairs.
- Sentiment Analysis Using Log-Generated Synthetic Text
Synthesized text from system logs enables sentiment prediction without frequent user surveys, reducing irritation and bias.
- LLM Tensor Network Compression for Resource Efficiency
Tensor network decomposition of LLM weight matrices compresses parameters, reducing energy consumption while maintaining model accuracy.
- Generalized Evolutionary Training Framework for Neural Networks
A generalized evolutionary training framework selects parent models and generates child models through parameter perturbation.
- DNN Training with Dynamic Zero-Reference RPU Crossbars
Dynamic zero-reference computation in RPU crossbar arrays reduces noise and hardware bias during deep neural network training.
- Neural Architecture Search Space Limitation for Comprehensible Models
A search space limitation apparatus constrains neural architecture discovery within predefined regular patterns to generate comprehensible model structures.
- Complex Iterated Least Square Thresholding for Neural Network Training
A complex iterated least square thresholding algorithm trains neural networks using real and simulated multipath signal data.
- Optimal Separable Convolutions for Deep Neural Network Efficiency
A separable convolution module calculates internal groups, channel size, and kernel size to optimize deep neural network processing.
- Autoencoder Text-to-Speech Synthesis Reducing Computational Load
An autoencoder generates fixed-size speech unit representations for natural-sounding text-to-speech synthesis.