Neural Architectures TRIZ Cases
- Custom Operator Decouples Candidate Data From Neural Network Graph
A custom operator inside a neural network graph structure acquires candidate datasets stored outside the graph.
- Multitasking Neural Network Segmentation for Independent Task Adaptation
A multitasking neural network architecture segments cross-task and task-specific layers to enable independent parameter updates.
- Modular Deep Learning Acoustic Model for Context Adaptation
A modular deep learning acoustic model uses shared layers and context-specific sub-modules to adapt speech recognition across varying environments.
- Gated Unit for Recurrent Neural Networks Using Element Processor
A gated unit processes hidden state vector elements using a dedicated memory array and element processor.
- Engagement Signal Classification for Collaboration Noise Reduction
Classification system processes raw collaboration signals into distinct engagement types using neural network models.
- AI Bias Prediction via Structured Test Data Generation
A system generates structured test data to predict AI model bias likelihood using payload logging analysis.
- Anomaly Detection Using Convolutional Auto-Encoder
Convolutional auto-encoder detects anomalies in time series data using unsupervised learning, adapting to changing patterns without manual rule configuration.
- Named Entity Recognition Using Spatial Feature Maps
A convolutional neural network estimates tag locations within a spatial feature map to recognize named entities in document images.
- Recurrent Neural Network Training Using Region of Target Updates
A recurrent neural network training module updates weights using a region of target to accelerate convergence rates.
- Recursive Array Layout for Transposable Synaptic Weight Access
A recursive array layout enables transposable access to synaptic weights, reducing memory accesses and power consumption in event-driven architectures.
- Multi-Horizon Time Series Forecasting with Attention and LSTM
A forecasting system combines sequence-to-sequence layers with temporal self-attention to process time series data.
- Autoencoder Anomaly Detection for Distributed Computing
Autoencoders encode user requests into latent space to generate normalcy scores, reducing false positives in distributed computing environments.
- Neural Network Explainability via Intermediary Logistic Regression Model
A data analysis apparatus uses an intermediary logistic regression model to replicate deep learning predictions for interpretable feature importance.
- Database-Driven Place and Route for CGRA Compilation
Iterative database-driven place and route system classifies unit graphs using a graph neural network to assign placement positions.
- Database-Driven Place and Route for Coarse-Grain Reconfigurable Architectures
A database-driven place and route system classifies unplaced unit graphs using a graph neural network to assign placement positions on configurable units arrays.
- Graph Data Interpretation via Alternative Sample Retrieval
A trained model acquires scores for graph data and retrieves alternative samples when initial values fall below a threshold.
- Object Detection via Attention Maps and Word Embeddings
Conditional inputs from a tagging network enable object detectors to generalize to unseen classes without scaling training datasets.
- Neuromorphic Error Corrector for Synapse Learning Accuracy
An error corrector in a neuromorphic device adjusts synapse resistance via post-synaptic feedback signals.
- Monotonic Neural Network Layers Using Composite Activation Functions
Composite activation functions combine convex and concave monotonic components to build flexible neural network layers.
- Batch Normalization Layer Training Using Pre-computed Statistics
A batch normalization layer training method applies pre-computed normalized statistic values to gradients.
- Dimensionality Reduction for Neural Network Descriptor Efficiency
Iterative dimensionality reduction extracts core descriptors from large sets to optimize neural network training workflows.
- Analog Neural Network Folded Circuit for Ultra-Low Power Wearable Sensors
A folded analog neural network circuit reuses layer circuitry across multiple cycles to process physiological data at the sensor node.
- Neural Network Connectivity via Event-Driven LFSR Computation
Linear feedback shift registers generate neural connectivity patterns from seed values, eliminating large matrix storage and lowering power consumption.
- Cognitive Pattern Templates for Engine Mix Settings
Cognitive pattern templates analyze racing scenarios to generate optimized engine mix settings.