Neural Architectures TRIZ Cases
- Thermodynamic Oscillator Chips for Faster Layer Normalization
Classical layer normalization can raise latency and energy use; thermodynamic oscillators compute mean, variance, and reciprocal operations through physical evolution.
- Thermodynamic SoftMax Gadget Using Engineered Oscillator Potentials
Engineered-potential oscillators compute SoftMax thermodynamically, avoiding classical conversion steps that add delay and measurement errors.
- Memristor Crossbar Weight Mapping for Neural Network Hardware
See how linked weight sets let memristor crossbars use positive hardware weights while software handles negative-weight flexibility.
- RC/RLC Neuron Circuits for Low-Power Spiking Neural Networks
Passive RC and RLC circuits implement neuron spiking, while sparse activity lowers power demands for scalable classification.
- Non-Linear Current Attenuation in Neural Networks for Synapse Drift
A non-linear current attenuator compensates for drift and dispersion in phase-change synapses, improving neural network operation.
- Spiking Neural Network Weight Updates Without Time Measurement
Time-difference-free synaptic updates use post-neuron potential to preserve inference accuracy when spiking inputs have low density.
- Autoencoder Restoration of Missing Time-Series Values with Limited Data
Missing energy-usage data is restored with an autoencoder trained on preprocessed time series and mean squared error, even with limited data.
- Global Style Token Embeddings for Data-Efficient Speaker Adaptation
Global style tokens generate multiple speaker embeddings and select a final match, reducing data and full-model fine-tuning needs.
- Domain Name Identification with Sequence Matrix Classification
Character-level comparison can miss malicious domains; sequence matrices and neural classification improve identification accuracy without a feature library.
- Multi-Scale Neural Network Distillation for Dense Predictions
Multi-scale backbone features, task prediction units, and distillation improve dense predictions such as depth maps and semantic images.
- Deep Learning Classifiers With Non-Volatile Memory Synapses
Non-volatile memory cells store tunable weights and compute in parallel, addressing bulky CMOS synapses and energy-intensive data movement.
- Neural Network Activation Circuits With Parallel Function Multiplexing
Software activation calculations consume time and power; parallel circuits and a multiplexor enable fast function selection.
- Neural Signal Detection Using Adverse Event Dependency Modeling
Traditional methods miss adverse-event signals or trigger false alerts; dependency modeling links report objects for more accurate pharmacovigilance detection.
- Mixed-Precision Quantizer Search for Lower-Cost ANN Training
EMS samples one quantizer bit-width alongside a full-precision pass, cutting training time and memory use for mixed-precision ANN search.
- Handwriting Recognition With Selective Language Model Refinement
A loss prediction module sends uncertain handwriting outputs to a language model, improving word accuracy while limiting computation.
- Neural Network Compression Through Calculator-Aware Operation Matching
Calculator allocation sets the reduced network's operation count, shortening processing time while keeping multiple calculators effectively utilized.
- Cookie-Space Domain Adaptation for Cookieless Device Attributes
Domain adaptation aligns cookie-based and cookieless request data so models predict device attributes consistently for content delivery.
- Neural Network Training With Selective High-Loss Processing
A batch-training approach filters loss values before gradient calculation, reducing neural-network operations and shortening time to convergence.
- Afferent Neuron Circuit With Volatile Threshold Switching
Complex CMOS afferent circuits limit scalability; a resistor and volatile threshold switch simplify integration while encoding stimulus strength as frequency.
- Tunable Gaussian Heterojunction Transistors for Compact Spiking Neurons
Dual-gated MoS2–carbon-nanotube heterojunctions tune Gaussian responses for single-transistor spiking neuron circuits.
- Neural Network SynFlow Pruning to Prevent Layer Collapse
Iterative SynFlow scoring prunes neural-network parameters before training, preserving synaptic flow to avoid layer collapse.
- Neural Network Execution Blocks for Covariate-Aware Forecasting
Separate covariate and target-variable models isolate external effects and improve forecasts when training data is limited.
- Hierarchical Neural Network Feedback for Noisy Sequence Recognition
Higher-level semantic feedback through apical synapses modulates lower regions, helping networks recognize sequences despite noisy inputs.