Physical Realisation TRIZ Cases
- Basis-Function Encoder Projection for Edge Neural Network Content Generation
Sparse basis-function projections reduce neural-network compute for edge content generation.
- Hardware-Specific Neural Network Masking for Cross-GPU Efficiency
Quantization, pruning, and masks adapt neural networks across GPUs without retraining.
- Optical Machine Learning System for Efficient Gradient Computation
An optical system uses spatial light modulation to generate synthetic gradients for electronic models.
- Spin Element Diffusion Barrier for Wiring Stability
An intermediate layer reduces elemental diffusion between the wiring and conductive parts, preventing material deterioration at connection points.
- Slot Attention Module for Entity-Centric Representation
A slot attention module initializes and updates entity vectors via key-query value functions.
- Distributed Neural Network Resilience via IoT Neuron Segmentation
Segmenting a neural network across multiple computing elements maintains model availability when individual devices fail.
- Pooling Calculation Device Multi-Channel Parallel Processing
The pooling calculation device matches channel count to internal memory bandwidth, easing pressure on neural network processing speed while maintaining accuracy.
- NPU Activation Pairing for Throughput and Power Efficiency
Pairing activations sharing common weights reduces NPU memory footprint and power consumption while maintaining high throughput.
- Programmable Activation Function Execution Unit for Neural Network Inference
A programmable activation function execution unit processes neural network data using piecewise linear approximations stored in lookup tables.
- Neural Network Quantization Optimization via Hardware-Aware Bit-Shift Configuration
Optimization constraints configure hardware bit-shifts to minimize quantization errors, resolving accuracy loss during model size reduction.
- Binary Neural Network Training via Shadow Weight Distributions
A binary neural network training method uses continuous shadow weight distributions to stabilize gradient descent and reduce rounding noise.
- Deep Neural Network Category Identification via Auto-Encoding
Integrates linear category analysis into deep neural networks for rapid data classification.
- Multi-GPU Neural Network Training via Stratified Data Sampling
Stratified data sampling balances GPU processing times by assigning similarly sized samples, reducing computational waste from execution jitter.
- Neural Network Compression via Reinforcement Learning
Reinforcement learning optimizes pruning and quantization parameters to resolve the contradiction between neural network performance and accuracy.
- Stacked Neuron Transistor Structure for High-Density Neuromorphic Computing
Vertical stacking of junctionless neuron transistors increases packing density while maintaining power efficiency and reliability.
- Optical Convolutional Computing Device Multi-Wavelength Interference Reduction
Multi-wavelength optical convolution device separates light into independent channels, eliminating interference and improving computational accuracy.