Machine Learning TRIZ Cases
- Black Box Model Prediction Characterization via Input Perturbation
A characterization system generates perturbed input data points to analyze machine learning model predictions and identify influential features.
- Deep Learning Model Conversion for Heterogeneous Hardware
Generates prediction models by adjusting parameters and operators based on topology descriptions, enabling heterogeneous chip acceleration.
- Adaptive Gaming Performance Score System
System dynamically adjusts gaming performance scores by correlating individual metrics, resolving skill mismatch frustration.
- Transfer Learning System for Deep Neural Network Model Selection
A transfer learning system selects pre-trained models using clustering algorithms and normalized mutual information.
- Decentralized Training Aggregating Disparate Feature Sets
Aggregates model updates from disparate client feature sets to calculate global weights, resolving uniformity constraints in federated learning.
- Simulated Delay Latency for Offline Online Feature Parity
System simulates online event processing delays during offline training to align feature distributions and reduce model refit requirements.
- Vector Modeling System for Distributed Data Transformation
A vector modeling system transforms structured data into machine learning vectors using tailored transformation strategies.
- Adversarial Learning Model with Confidence-Weighted Loss Filter
A learning unit filters data using correct answer labels to predict adversarial example labels.
- Distributed Voting Mechanism for Low Power Lossy Network Attack Detection
Segmented voting domains and dynamic thresholds reduce false positives and network overhead while maintaining high detection reliability.
- Machine Learning Parameter Optimization via Discontinuity Detection
A learning apparatus segments parameter spaces to identify discontinuity points and calculate estimated performance improvements.
- ML Skew Prediction for Multi-Tier Storage Data Movement
Predictive models convert skew measurements across time granularities to optimize data migration while reducing computational overhead from real-time analysis.
- Just in Time Classifier Training for Missing Data
A classifier calibration method modifies training datasets to match test data features.
- Database Machine Learning Model Deployment via Integrated Functions
Integrates trained model data as a database function to eliminate external data export and reduce processing costs.
- Conditional Normalizing Flow for Posterior Collapse in CVAE
Flexible conditional priors resolve posterior collapse, enabling accurate multi-modal traffic prediction.
- Normalizing Flows for Privacy-Preserving AUC Estimation
A normalizing flows model estimates probability density functions to compute the area under the receiver operating characteristic curve.
- Adaptive SD-WAN Policies via Machine Learning
Site-specific local policies generated by machine learning optimize network path selection and traffic steering for unique conditions.
- Heterogeneous Scheduler for Neural Network Execution
A heterogeneous scheduler dispatches neural network operations across CPU, GPU, and neural processors using weighted graph traversals.
- Dynamic Web Page Load Mitigation via Threshold-Based Alternative Content
A web application system detects slow page loads and serves alternative content to maintain user engagement.
- Self-Supervised Counterfactual Transformer for Temporal Confounding
SCOT framework segments treatment estimation via temporal and feature-wise attention, addressing temporal confounding bias while reducing computational runtime.
- Knowledge Graph Modeling for Chemical Data Automation
A computer system trains machine learning models using knowledge graphs to predict component attributes in production environments.
- Machine Learning Predictor Profiling via Stratified Sampling
Stratified sampling determines predictor intervals to build accurate local models, resolving the trade-off between explanation precision and sampling time.
- Feature Map Decomposition for Streaming Data Latency Reduction
Segmenting processing into overlapping sliding windows generates partial feature maps, reducing latency while maintaining accuracy.
- Generalized Balanced Few-Shot Learning Without Base Data
Pre-trains feature extractor then fine-tunes joint linear classifier via weight normalization to balance base and novel class performance without base data.
- Classifier Committee Weighting by Annotator Feature Rankings
A committee of classifiers weights models based on annotator-provided feature contribution rankings to generate prediction labels.
- Dynamic Computational Resource Allocation for ML Models
A dynamic allocation system adjusts computational resources based on real-time model requirements.
- Machine Learning Transaction Settlement Rule Generation
An AI system analyzes historical deduction-promotion pairs to generate settlement rules, resolving manual processing bottlenecks in B2B transactions.
- Machine Learning User Presence Estimation for Power Mode Control
A machine learning model estimates user presence using time-series sensor data to optimize information processing apparatus operations.
- Container Orchestrator Tuning via Workload Persona Profiling
Automated workload profiling identifies container personas to map precise orchestration policies, resolving manual tuning bottlenecks.
- Neural Network Processing Subdivision for Stride Adaptation
Subdividing input data arrays into portions enables multiple processing passes to handle varied neural network strides without hardware changes.
- Local Machine Learning Model Fine-Tuning for Privacy Preservation
A local machine learning model fine-tunes pretrained weights using user-specific data on the device.
- THz Network Supervisor for Transient Device Data Management
A supervisor manages permanent and transient devices in THz zones to enable data deduplication and compression.
- Dynamic Composite Data Structures Using Feature Envelopes
Function mappers map extracted feature envelopes to synthesis parameters, preserving natural correlations while reducing dimensionality.