Machine Learning TRIZ Cases

Adapting subscriber-agnostic ML classifiers with tenant data improves sensitive data detection, governance accuracy, and compliance control.

An ASIC k-NN architecture uses memory, distance calculation, and sorting blocks to cut AI power and cost without CPUs or GPUs.

A trained ML model scores and ranks biological workflow outputs, cutting manual review time while preserving output selection accuracy.

A page-based natural language setup flow lets users create and release customized digital assistants without coding or complex configuration.

Unified embeddings from image, video, audio, and text improve video risk classification while reducing handcrafted logic for new categories.

Historical speaking-time data and adaptive scaling help estimate topic allotments more accurately and reduce conference scheduling errors.

Global model identifiers unify registration, conversion, and execution across network and terminal devices to prevent AI model incompatibility.

Combining Doppler velocity, RCS, and geometric distance improves point cloud similarity evaluation for generative model training in ADAS.

Targeted retraining by skin tone, lighting, and hue helps machine learning models reduce bias and improve accuracy for underrepresented groups.

Local edge training and partial cloud model updates cut data transfer and server load while preserving cloud service accuracy.

NLP-guided model updates and automated threshold setting improve time series anomaly detection while reducing false positives from seasonality.

Configurable learning blocks let UEs and base stations train selected DNN layers, cutting training and signaling overhead in wireless networks.

Machine learning reads TLS handshake sequences to identify OS type and version more reliably than manual rules, even without explicit identifiers.

A dynamic margin adjusts to embedding hardness during XMC training, improving label separation, inference accuracy, and resource use.

Feasible multi-step recourse paths use distance thresholds and transition labels to turn negative ML decisions into actionable positive outcomes.

Preconfigured AI/ML model IDs and settings let user equipment switch configurations during base-station handover with less signaling delay.

Uncertainty-guided parallelism allocates compute by sequence segment to improve selective state space model efficiency and prediction reliability.

Preconfigured target cells and early synchronization cut 5G handover signaling overhead, latency, and data loss during cell switching.

On-chip fine-tuning offsets dense memory noise in stored ML models, preserving inference performance on resource-limited computing architectures.

User location, device, and network state are checked before payment approval, triggering extra authentication when confidence is low.

A routing decision model uses preconfigured network data to choose secure real-time payment paths across multiple transaction networks.

Class-priority sampling focuses annotation on low-accuracy classes, cutting labeling cost while improving multi-class classifier accuracy.

Weighted ensemble models and similar-product matching improve demand forecasts while letting users adjust model emphasis to reflect intent.

Event-triggered E2SM REPORT and QUERY let near-RT RIC keep UE context current across split CU-DU RAN for non-GoB beamforming.

Automated feature extraction converts time-series signals into static model inputs, reducing manual engineering and compute needs for forecasting and imputation.

Trained models flag sensor-based compliance deviations, score their impact, and deploy updated protocols to cut resource use and avoid obsolescence.

Separate management and task links let a client route AI workloads to the best local or networked device for secure, efficient reasoning.

Automatically generated AI prompts turn incident data into claim summaries, cutting manual processing delays while preserving structured accuracy.

Converted features align heterogeneous data across organizations, enabling federated prediction of intervention effects without sharing raw data.

Selected schema and language-matched pool entries give an LLM richer context to generate complex query logic with lower processing overhead.

Visual signals flag unsupported or risky generated text so users can revise model outputs faster while improving factual trust and source checking.

Proactive memory recovery uses compression swapping and thread selection to load large AI models without volatile memory saturation.

Network-side ML combines RAN, core, traffic, and geographic data to locate edge devices without GPS, cutting power use and improving reliability.

Estimated ink load and detected tension let later print sections correct color misregistration caused by paper stretch during continuous printing.

Fine-tuning target and nuisance tasks removes undesired bias from self-supervised models, improving speech emotion recognition with limited labels.

Continuous metric monitoring lets 5G nodes detect degraded AI radio models, trigger retraining, and maintain reliable radio function performance.

Multiple gating layers are trained jointly to avoid overfitting in feature selection while reducing neural model compute and memory use.

Automated anomaly detection and AI-generated adaptation proposals speed digital artifact review while preserving contextual judgment.

Clustering unlabeled samples adds new features to labeled data, improving model generalization when labeled training data is scarce.

Multiple AI devices use speech volume, sensor feedback, and intention analysis to identify the right responder and target device.

Context analysis triggers voice assistance only when reading is difficult, improving accessibility while avoiding always-on energy use.

Browse shelves act as intermediaries to map search queries to relevant facets, improving filtering accuracy without overwhelming users.

Synthetic minority time series and selective down-sampling rebalance rare-event training data to improve classifier accuracy.

Sample and feature alignment across 5GC network elements enables vertical federated learning with lower coordination overhead for analytics services.

AI models detect flavor shift paths and combine targeted corrections to preserve characteristic food flavor during processing and storage.

Machine learning fuses RAN, core, traffic, and geographic data to locate edge devices in real time without GPS power and reliability limits.

Synthetic time-series generation uses constraint learning and diffusion decoding to expand dynamic-system training data without costly extended measurement.

Unsupervised merchant and user grouping enables secure card registration across selected accounts while reducing repeated entry and theft risk.

Distributed AI/ML processing via NWDAF and edge servers cuts 5G delay while preserving synchronized end-to-end inference across PDU sessions.

An LLM analyzes user input and on-screen content to generate UI suggestions that cut search time in unorganized applications.