Neural Architectures TRIZ Cases
- Power-of-2 Parametric Activation Clipping for CNN Quantization
PACT2 function clips activation feature maps to power-of-2 ranges, resolving 8-bit inference accuracy loss in embedded devices.
- Spiking Neural Network Back-Propagation Using Binary Error Encoding
A spiking neural network back-propagation algorithm uses binary or ternary error encoding to adapt hardware constraints.
- Neural Network Parameter Quantization for Memory Reduction
A neural network parameter analysis method reduces memory size by evaluating quality factors across modified precision levels.
- Series-Connected MTJ Stacks for Spiking Neural Network Neuron Circuits
Series-connected magnetic tunnel junction stacks driven by pulsed power supplies increase spike rates and control variability in spiking neural networks.
- Neural Network Quantization With Fixed Weight Range
Fixed weight range constraints stabilize neural network training while pruning weights to lower hardware complexity without sacrificing prediction accuracy.
- Neural Network Hardware Recognition Circuit
Hardware neural network processing replaces software with time signals and threshold value processing to reduce power consumption.
- Lambda Functions for Reduced Memory Complexity in Long-Range Interaction Modeling
Lambda functions transform context data into compact parameters applied directly to queries, enabling efficient long-range interaction modeling.
- Machine Learning Method for Sequence Prediction Using Sub-Sequence Weighting
A machine learning method generates sub-sequences from a target sequence to determine weights for prediction.
- Entropy-Based Logit Transformation for Neural Network Confidence Calibration
An entropy-based logit transformation function calibrates classification network confidence scores using learnable parameters.
- Self-Modifying Code for Neural Network Computation Reduction
Self-modifying code overwrites redundant neural network instructions with NOPs, eliminating unnecessary computations when break conditions are met.
- Brain-like Visual Neural Network Forward Learning
A brain-like visual neural network uses forward learning and meta-learning to encode position information efficiently.
- Sub-matrix Input for Neural Network Keyword Detection
Segmenting input matrices into sub-matrices allows neural network nodes to detect keywords with low latency while maintaining accuracy.
- Neural Network Pruning via Self-Distillation and Weight Reintroduction
Iterative weight reintroduction combined with knowledge distillation removes neural network parameters while maintaining accuracy, reducing memory footprint.
- Neural Network Parallelization Plan Generation
Computing system generates parallelization plans for neural networks using virtual tensors to track dependencies and schedule fine-grained tasks across multiple devices.
- Dynamic Synapse Type Conversion for Neural Network Resource Management
A method converts plastic synapses to fixed types based on training status to optimize hardware resource allocation.
- Federated Neural Network Training with Local Sample Distribution
A federated learning system reconstructs neural network models using local sample distribution data to update submodels on member devices.
- Tunable Waveguide Phase Shifters for Photonic Neural Networks
Adjustable waveguide widths in Mach-Zehnder interferometers configure optical phase shifts without active energy consumption.
- Adaptive Token Depth Adjustment in Transformer Networks
Dynamic halting scores skip unnecessary layers for low-importance tokens, balancing accuracy with reduced latency and power consumption.
- Image Sensor Neural Network Processing Structure
Neural network processing structure directly connects photo detectors to eliminate analog-to-digital converters and random access memory.