Neural Architectures TRIZ Cases
- Sparse Neural Network Inference With Thresholded Weights And Outputs
Thresholded sparse weights and outputs cut computation and memory while preserving accuracy.
- Neuromorphic Unit With Orthogonal Spikes for Parallel Workloads
CDMA-based orthogonal spike modulation combines workload streams while preserving payload capacity for parallel neural processing.
- Hierarchical Mode-Wise Pooling for Accurate Feature Vector Processing
Pooling, transformation, and unpooling preserve hierarchical indices while improving accuracy and reducing training time.
- Meta-Gradient Options for Stable Multi-Task Reinforcement Learning
A manager network and option policies discover reusable action sequences, reducing data and computation demands across tasks.
- Octree-Based Convolutional Neural Networks for Sparse 3D Shape Processing
An octree stores empty regions as leaf nodes, enabling GPU-ready convolution with lower storage and computational demands.
- Memristor-Based Neuron Simulation Circuit for Spatiotemporal Processing
This case uses dynamic and threshold-switching memristors in neuron circuits to emulate biological dynamics for spatiotemporal tasks.
- Binary Neural Network Channel Pruning by Weight Flipping Frequency
Weight flipping frequency guides structured BNN channel pruning, delivering 20-40% fewer binary operations with limited accuracy loss.
- Machine Learning Opportunity Cost Prediction Without Demand Forecasting
A machine learning model uses capacity, expiration time, and historical transactions to improve pricing predictions with insufficient data.
- AI Image Recognition with Letter-Sequence Prediction for Unique Names
A nodal neural network predicts missing letters from recognized ones, improving recognition of unique handwritten names.
- Self-Attention Battery Cycle Life Prediction With Physics Tuning
A physics model tunes a self-attention learner with ground-truth loss curves and cycle data to improve battery life predictions.
- Neural Network Detection Model Testing with Error Threshold Feedback
This case uses hidden-layer error checks and signal reacquisition to limit unnecessary computation while improving detection accuracy.