Neural Architectures TRIZ Cases

A topological stochastic model extracts and aggregates feature vectors to verify electronic signatures.

A CNN accelerator reuses intermediate convolved data values within on-chip buffers to minimize off-chip transfers.

Automated neural architecture search uses evolutionary algorithms to optimize design parameters, reducing manual effort and computational costs.

Dual mesh network prevents cyclic deadlocks by advancing time steps only after all spikes are received.

Front-end cores fetch sparse input feature map and kernel tiles to perform convolutional operations without computing zero values.

Segmenting architectures into reusable blocks reduces time and computational resources while improving inference accuracy by 6.7 percent.

Multiple discriminators apply distinct deformation processing to input data, enabling coordinated updates that stabilize generative adversarial network training.

Clustering-based quantization compresses neural network weight tensors by separating outliers from inliers and applying vector quantization.

Cascaded directional couplers and modulating elements enable optical associative learning operations.

A computing device implements multi-synapse projections with distinct synaptic time constants to process spike trains.

Synapse circuits supply charge to soma circuits using resistor elements, enabling neuron potential monitoring without dedicated hardware.

Merged weight matrices and instruction packages reduce DMA overhead and latency by enabling continuous data access without frequent memory reloads.

Continuous feedforward sinusoids adjust neuron positions and weights to reduce training time and computational resources compared to back propagation.

A stochastic transmitter algorithm converts inputs into data symbols using trainable neural network parameters.

Iterative pruning and splicing reduce memory and computation requirements by approximately 18 to 100 times without sacrificing accuracy.

A neural network dynamics model reconstructs spiking data from local field potential inputs.

Replacing exponential decay with a step function reduces processing load and power consumption while maintaining recognition accuracy.

Local memory caches neural network models to bypass host bandwidth limits and reduce model loading latency.

A convolutional neural network dilates sparse point observations to generate dense raster grids for machine learning inputs.

Simultaneous knowledge distillation and sparse pruning reduce neural network size without destroying fine-tuned model knowledge.

A neural network classification device processes image and numerical data through a heterogeneous integration module, resolving IoT model training limitations.

Segmenting neural network processing between vehicles and edge servers reduces data traffic and battery consumption while maintaining high reliability.

A robust learning device selects a subset of neural networks to update parameters and minimize prediction loss.

Machine learning model extracts content features to determine personalized initial bid costs, resolving inaccuracy from static formulas.

Segmenting convolutional filters into clusters and calculating local geometric medians reduces inference latency while preserving model accuracy.

A Taylor series activation function unit computes neural network approximations using multipliers and accumulators.

Pre-normalized layers compute statistics locally to remove batch-level synchronization overhead.

Augmented reality generates synthetic visual inspection datasets by creating and manipulating 3D models of anchor objects for computer vision applications.

An optimization device adds a coefficient-based offset to energy changes during state transitions.

A fall prediction model uses an ensemble of recurrent neural networks to process diverse data streams.

Automatic weighted deep fusion architecture integrates multi-modal features through embedding and discriminative levels.

A memory device integrates external and internal accelerators to distribute neural network operations across hierarchical processing units.

A student neural network mimics teacher model outputs to compress architecture while preserving recognition accuracy.

A machine learning system trains models using sparse data through iterative Monte Carlo sampling and convergence testing.

A neural network memory system adapts query recirculation steps via a reinforcement learning controller to retrieve knowledge items.

BLSTM peak detection segments speech files, stabilizing prediction results while reducing training time.

A graph neural network apparatus separates user-item interactions into positive and negative graphs to embed nodes in a common virtual space.

A denoising convolutional auto-encoder enhances fingerprint images using synthetic training data.

A neural network learning device adjusts activation function parameters to converge toward linear forms for efficient weight aggregation.

Estimate probability distributions of normal operational data to detect anomalies without labeled samples, resolving high dimensionality and noise challenges.

A neural network calculates internal state and perturbation values simultaneously during forward propagation to enable continuous online learning.

PatDNN uses pattern-based pruning and compiler optimizations to resolve the trade-off between model compression ratio and processing throughput.

Segmenting input feature data into subdata blocks reduces data transportation times by buffering only active segments, resolving cache capacity constraints.

Layered quantization reduces storage space while maintaining calculation accuracy in mobile AI applications.

A selection prediction model estimates shopper order processing duration for online concierge systems.

A hybrid training method switches neural network learning algorithms to balance convergence speed and generalization performance.

Interleaved tiling reduces rotation and multiplication overhead in fully homomorphic encryption convolution operations.

Nested sub-networks adapt to resource constraints without retraining, resolving batch normalization inconsistencies.

Electronic device dynamically adjusts neural network weight matrices based on real-time resource information.