Neural Architectures TRIZ Cases

Compare RNN activity across time windows to reduce noise and adversarial instability.

This case segments spike generation and uses uniform random selection to preserve biological compatibility with less hardware complexity.

Coal quality and operating data train an Adam-optimized BPNN to predict flue dust emissions and reduce manual CEMS verification.

Local model exemplars use federated averaging and constrained clustering to detect anomalies while protecting private edge data.

Reordering weights into sparse bitmaps and coefficient groups reduces storage and transmission needs while lowering inference power use.

Node-group allocation adapts to activation patterns, cutting execution time and power.

Segmenting large minibatches into smaller spatial units reduces off-chip data fetching and eliminates waiting time between processing layers.

A learning apparatus trains machine learning models with reduced sizes by setting specific training conditions to optimize performance.

Pre-train LSTM layers on unlabeled data to reduce training time and computational resources while improving classification accuracy.

Slave devices train local model versions using private data and send gradients to a master device for global updates.

Segmenting offline training from online inference reduces computational complexity, enabling accurate time-series classification on constrained edge hardware.

A vector neural network architecture calculates output vectors using prediction matrices to eliminate iterative dynamic routing.

Feature engineering extracts structured vectors from non-structured IoT data streams for neural network classification.

Modified error function combines empirical loss with Vapnik-Chervonenkis dimension bounds to control model complexity and improve generalization.

A neural network system condenses mobile app communications by clustering topics and removing duplicates to generate personalized user options.

A neuromorphic memory device with a crossbar array structure performs in-memory multiply-accumulate operations using incremental state changes.

An adaptive audio-generation model learns new speaker embedding vectors to emulate voices with minimal training data.

A Bayesian Spatiotemporal Graph Transformer network predicts multi-aircraft trajectories using deep learning.

A spiking neuron encoding module generates sparse spike train representations from continuous input signals using convolution kernels with time-varying thresholds.

A computing graph compilation method supports multiple variable input ranges through automatic runtime selection of optimized graphs.

Deep CNN with bag-of-words reconstruction enhances 3D facet description accuracy.

Separates frozen and trainable neural network layers into distinct chiplets with enhanced memory, reducing transition time between training and deployment.

A learning content recommendation system calculates an expected score to determine optimal questions for users.

Random probing generates a synthetic dataset to train a student model that replicates target behavior, preserving privacy and avoiding full retraining.

Virtual chips in a decentralized network validate AI model performance, reducing computational burden on individual devices.

A neural network growth method selects connections based on weight thresholds to reduce size and computational requirements.

A neural network training method aligns class activation maps between original and adversarial data using logit pairing constraints.

Early growth selectively expands channel sizes in intermediate layers, recovering accuracy while resisting bit-flip attacks.

A solving method converts problem definitions into constraint expressions and propagates implications to identify logical contradictions.

A neural network fuses sensor signals using sparse stochastic neurons that reduce computing operations by zeroing low precision values.

Dynamic soft mask pruning adjusts weight values over training epochs to control sparsity and pace, preventing accuracy degradation from hard masking.

A neuromorphic neural network architecture using distributed binary representation to abstract sensory data without optimization.

Deep neural networks translate spoken natural language into structured vehicle commands, resolving sensor data limitations for safer navigation.

System searches candidate normalization-activation layer architectures for neural networks to identify optimal designs.

A hierarchical neural network architecture uses parallel first and second networks to calculate feature data via specific channel portions.

An NCE decoder reduces popularity bias in a two-headed attention fused autoencoder by increasing likelihoods for items with observed interactions.

Graph external attention module integrates external view information with internal graph features to enhance node representation accuracy.

Expanding input feature maps in a multi-layer neural network model increases weight connections for higher precision without requiring additional storage space.

A second neural network generates task-specific parameters to modify a first neural network structure.

A neuron circuit performs online learning by integrating synaptic weight updates directly into the hardware structure.

A pulse driving apparatus adjusts synaptic weights using controlled voltage pulses to minimize asymmetry between long-term potentiation and depression processes.

Progress scores in a recurrent neural network trigger output emissions before full input processing, reducing latency and computational resources.

Adversarial training of generator and discriminator networks improves local search diversity to escape deep learning local extrema.

A system extracts video fragments as stickers by identifying facial emotions through pre-trained convolutional neural networks.

A spiking neuron apparatus uses a coherent Ising machine to simulate neural dynamics via optical pulses.

RT-TCN algorithm reuses prior convolution outputs, eliminating redundant computations to resolve excessive resource consumption during real-time evaluation.

A computing apparatus selects a backbone architecture and calculates weights for candidate operation blocks to configure an optimized neural network structure.

A neural network analyzes temporal communication images to derive precise device fingerprints and reverse-predict MAC addresses.

Segmenting recurrent and network computational models reduces training time and memory usage while resolving numerical stability issues in dynamic environments.