Neural Architectures TRIZ Cases
- Partially Shared Multi-Task DNNs for Scalable Learning
Constraint-based sharing lets multiple DNNs reuse common layers while retaining task-specific layers, addressing non-convex convergence issues.
- Superpixel CNN Layers for Efficient Shallow-Depth Computation
Grouping spatial inputs into deeper superpixels makes shallow-depth CNN matrix multiplication faster without hardware changes.
- Multi-Dimensional Machine Learning for Resource Value Prediction
Conventional resource management can misjudge resource and activity values; multi-dimensional ML predicts both to trigger timely automated actions.
- Sparse Multi-Layer Model Explanations Using Latent Activation States
Limited latent-feature connections and activation-state matrices make complex neural-network outputs easier to explain deterministically.
- Parallel Neural Network Finetuning With Block-Diagonal Matrices
Block-diagonal matrix multiplication lets neural network finetuning run in parallel while reducing computation time and resource consumption.
- MRAM Neural Network Platform With STDP Learning for Parallel Processing
STDP adjusts MRAM synaptic weights for parallel pattern recognition despite von Neumann bandwidth limits.
- Compound Coefficient Scaling for Resource-Aware Neural Networks
Resource-aware scaling assigns computational capacity across width, depth, and resolution to improve accuracy while controlling training and inference cost.
- Block-Diagonal Neural Network Finetuning for Parallel Computation
Block-diagonal matrices split neural-network finetuning across cores, reducing computational overhead while preserving model performance.
- Two-Pass Speech Recognition With Deliberation for Low Latency
See how a deliberation network combines first-pass RNN-T hypotheses with acoustics to reduce WER while meeting on-device latency constraints.
- Physics-Augmented Neural Networks for Mixed Order-Chaos Dynamics
Generic-coordinate preprocessing and Hamiltonian loss help forecast nonlinear dynamics while mapping transitions between orderly and chaotic states.
- Adversarial Information Bottleneck for Efficient Model Training
A discriminator–task-model minimax game estimates mutual information tractably, improving training efficiency and limiting overfitting.
- RNN Number Format Selection for Hardware Accelerator Execution
A test RNN collects value statistics to select a common number format, helping hardware accelerators execute recurrent networks with lower memory bandwidth.
- Online Time-Series Forecasting With Spiking Reservoirs for Edge Devices
Spiking reservoirs encode temporal spikes and use FORCE learning to reduce memory and computation for edge forecasting.
- Automatic Feature Grouping for Higher-Order Recommendation Accuracy
Complex feature interactions can slow training and reduce prediction accuracy; automatic grouping selects important higher-order interactions for recommendation.
- Spike Neural Network Circuit for Parasitic Charge Correction
A potential correction circuit and transistor limit parasitic charge entering the membrane capacitor, improving SNN charge-accumulation accuracy.