Neural Architectures TRIZ Cases
- Self-Training GANs for Semi-Supervised Learning
A self-training method augments generative adversarial network datasets by assigning pseudo-labels to generated samples.
- Multi-task Fusion Neural Network Architecture for Mobile Image Processing
A multi-task fusion neural network architecture uses a common feature extractor to process multiple image tasks concurrently.
- Deep Neural Network Parameter Tuning for Domain Adaptation
A deep neural network adjusts lower-layer parameters to align target-domain data distributions with training-domain spaces.
- Augmented Neural Network Sequence Processing for Output Accuracy
Sequence processing subsystem determines whether to generate additional neural network outputs based on initial results.
- Modular Deep Multitask Learning Framework for Cross-Domain Knowledge Sharing
Decomposing monolithic feature extractors into reusable pseudo-tasks resolves the trade-off between extraction precision and cross-task adaptability.
- Spiking Neural Network Device With Independent Synaptic Depressor
An independent synaptic depressor updates weights on a schedule, enabling low spike density learning without hardware complexity.
- Neural Network Generation via Differentiable Architecture Search
Differentiable architecture search optimizes neural network structures for efficient processing.
- Self-tuning DNN Compression for Resource-Limited Devices
A self-tuning model compression methodology reconfigures deep neural networks by pruning redundant neurons and quantizing weights.
- Neuromorphic Synapse Weight Update via Delayed Pulse Sequences
Delayed pulse sequences update synapse weights in neuromorphic devices, resolving slow speeds and stability issues from synchronized timing cycles.
- Sparse Finetuning Hardware Accelerator for Neural Network Weight Updates
Segmenting weights into fixed non-volatile and finetuned volatile memory enables fast context switching while reducing power consumption.
- Hybrid Parallelism for Neural Network Training Efficiency
Segmenting models across workers reduces synchronization overhead, overcoming scaling inefficiencies in large dataset training.
- Hierarchical Neural Network for Temporal Pattern Recognition
A hierarchical neural system processes input streams into population codes to cluster temporal patterns using Hebbian learning rules.
- Neuromorphic Circuit Routing via Gating Pulses
Timed gating pulses enable exact information transfer between neuronal populations by dynamically routing graded signals through controlled inhibition.
- Neural Network Quantization Method for Reducing Storage and Bandwidth
A neural network quantization method scales weights and input data to low-bit representations, reducing storage space and processing bandwidth.
- Sub-tensor Column Execution for Neural Network Cache Optimization
Segmenting neural network execution into asynchronous sub-tensor columns reduces memory bandwidth consumption and redundant calculations.
- Meta-Level Continual Learning for One-Shot Classification
Segmenting weights into fast and slow components resolves the trade-off between classification accuracy and continuous learning ability.