Menu
Skills in this repository
SkillsMP has collected 4,114 skills from hiyenwong/ai_collection. Open a skill to review its source and details.
hiyenwong/ai_collectionShowing 40 of 4,114 collected skills.
Symmetry-Protected Lyapunov Neutral Modes in Equivariant Recurrent Networks. Theoretical framework for when continuous attractors (zero Lyapunov exponents) are guaranteed by symmetry rather than tuning. Activation: equivariant rnn, symmetry-protected modes,…
DeepStack methodology for design space exploration (DSE) in system-hardware co-design. Scalable and accurate performance modeling for distributed 3D-stacked AI systems. Use when performing early-stage system design optimization, hardware-software co-design,…
Automated threat modeling and security verification for cyber-physical systems (CPS) and AI-enabled systems. Use when: (1) Designing security architecture for CPS/IoT systems, (2) Performing automated threat modeling from system architecture models (SysML,…
Tensor Network Feature Engineering methodology for multi-class neurological disorder prediction from MRI data. Uses tensor network decompositions to extract high-dimensional features from sparse medical imaging. Activation: tensor network MRI, neurological…
Framework for implementing Test-Time Training (TTT) - updating model weights during inference to adapt to continuous information streams. Use when designing adaptive AI systems, online learning for LLMs, breaking static train-deploy paradigm, or handling…
Perceptual misalignment of texture representations in convolutional neural networks — finds no connection between CNN Brain-Score and alignment with human texture perception, suggesting texture perception involves mechanisms distinct from object recognition…
Thermodynamic Networks methodology for autonomous physics-based computation using non-equilibrium steady states. Identifies Negative Differential Conductance (NDC) as the critical property for computational expressivity. Applies to quantum dot networks,…
"Training-free looped transformers — inference-time technique that loops a contiguous mid-stack block of layers in a frozen pretrained LLM to improve reasoning without fine-tuning. Use when: (1) improving a frozen LLM's reasoning at inference time, (2)…
Multi-Modal Graph Neural Network with Transformer-Guided Adaptive Diffusion for Preclinical Alzheimer Classification. Combines diffusion kernel (short-range) + multi-head attention (long-range) for brain network analysis. Activation: Alzheimer, preclinical…
Use prototype-based readout layers for transformer encoders to replace pooling methods (mean pooling, class token). Avoid information collapse with learned compression mechanism. Activation: prototype readout, transformer output layer, collapse-free…
Multi-Stage Warm-Start Deep Learning Framework for Unit Commitment. Transformer-based architecture for predicting generator commitment schedules with deterministic post-processing for physical feasibility, warm-start strategy for MILP solver, and…
End-to-end pipeline for large-scale portfolio selection with cardinality constraints using trapped-ion quantum computers. Use when: executing portfolio optimization on trapped-ion QPU hardware; solving QUBO subproblems via BF-DCQO; decomposing large…
TRIBE v2 tri-modal foundation model methodology for in-silico neuroscience. Uses video/audio/language embeddings to predict whole-brain fMRI across 720 subjects.
TRIBE v2 tri-modal foundation model methodology for in-silico neuroscience. Uses video, audio, and language modalities to predict human brain activity across naturalistic and experimental conditions. Supersedes linear encoding models with several-fold…
Triple-Loop Consolidation methodology for persistent memory in non-gradient dissipative cognitive architectures. Deep Memory (DM) operates through recording-seeding-reentry cycle. Discrete MoE routing is causally prerequisite. Activation: triple-loop…
Five-principle framework for building and governing trustworthy AI agents. Covers human control, value alignment, secure interactions, transparency, and privacy in agent architecture design.
Optimize tool/schema definitions for LLM agent deployments using TSCG principles. Converts JSON tool schemas into token-efficient structured text formats that small models (4B-14B parameters) can reliably interpret. Use when: (1) agent tool-use accuracy drops…
Turn-level Truncated On-Policy Distillation (TT-OPD) methodology for training multi-turn medical AI agents via reinforcement learning. Addresses multi-turn collapse, response length explosion, and tool-use erosion in clinical dialogue. Use when training…
Expert guidance for globally shared expert pool Mixture-of-Experts architecture. Based on UniPool paper (arXiv:2605.06665). Use when designing MoE architectures, expert pooling, pool-level balancing, NormRouter, sublinear expert parameter scaling, or…
Systematic RSA comparison showing untrained CNNs match backpropagation-trained CNNs in V1 visual cortex alignment. Large-scale fMRI analysis reveals that random feature detectors can capture V1 representational structure. Keywords: untrained CNN, V1 cortex,…
系统RSA比较研究：展示未训练CNN在V1视觉皮层区域与反向传播训练的CNN具有相似表征。通过大规模fMRI和表征相似性分析，挑战传统深度学习需要大量训练的观点。适用于视觉皮层建模、CNN可解释性、神经科学。
Source text: Chinese
Systematic RSA comparison showing untrained CNNs match backpropagation-trained networks at V1 visual cortex, revealing architecture's dominant role over learning rules in neural alignment. Activation triggers: untrained cnn, backpropagation, v1, rsa,…
Systematic RSA comparison showing untrained CNNs match backpropagation-trained CNNs at V1 visual cortex. Trigger words: untrained CNN, backpropagation, RSA, V1, representational similarity
Systematic RSA comparison showing untrained CNNs match backpropagation at V1 alignment with human fMRI. Evaluates BP, FA, PC, and STDP learning rules against THINGS-fMRI dataset using 720 stimuli across 3 subjects. Use when studying brain-model alignment,…
Systematic RSA comparison showing untrained CNNs match backpropagation-trained CNNs at V1 visual cortex. Untrained random-weights CNN (rho=0.076) exceeds backprop (rho=0.034) at V1/V2 (p<0.001). STDP achieves highest V1 alignment among trained rules…
Validation-driven LLM workflow pattern - using verification loops to ensure LLM-generated outputs are correct. Extracted from 'Generating Statistical Charts with Validation-Driven LLM Workflows' (arXiv 2026-05-01). Applicable to code generation, data…
Vector Policy Optimization (VPO) methodology for training LLMs to maintain response diversity during inference-time search. Use when: (1) Training models for test-time compute scaling, (2) Improving AlphaEvolve/AlphaCode-style search procedures, (3)…
VENCircuit methodology — Von Economo neurons as residual gradient scaffolds in recurrent spiking neural networks for reliable social skill acquisition. Use when researching: Von Economo neurons (VENs), spiking neural networks for social cognition, gradient…
Voice AI interaction patterns for building realtime voice agents with reasoning, tool use, and multilingual support. Based on OpenAI's GPT-Realtime model release patterns.
Wavelet-Enhanced Mixture-of-Experts (WaveMoE) foundation model for time series forecasting. Use when building time series prediction models, incorporating frequency-domain information, or designing MoE architectures for temporal data.
X-Token methodology from arXiv:2605.21699 (May 2026). Projection-guided cross-tokenizer knowledge distillation: P-KL (sparse projection matrix) and H-KL (hybrid relaxed matching) for teaching LLMs with incompatible vocabularies. Use when: cross-tokenizer…
Privacy-enhanced zero-order federated learning using multi-key homomorphic encryption (xMK-CKKS) over wireless channels. Methodology for secure FL aggregation without channel estimation, supporting N-1 client compromise tolerance. Use when designing…
Coarse feedback for human-aligned visual representations. Demonstrates that extremely coarse classification signals (e.g., 8 classes) produce representations that match or exceed brain alignment of fine-grained (1000-class) or self-supervised models. Use…
Implement early stopping strategies for reasoning models based on confidence dynamics. Detects when reasoning trajectories reach high confidence early vs. unproductive long reasoning traces. Use for: (1) optimizing LLM reasoning tasks, (2) reducing…
FADE: Forgetting through Adaptive Decay for continual learning. Adapts per-parameter weight decay rates online via meta-gradient descent. Balances acquiring new knowledge with retaining old. Activation: FADE, adaptive weight decay, continual learning…
Fast efficient coding and sensory adaptation in gain-adaptive recurrent networks — unified mechanistic model reconciling adapter-repulsion and prior-attraction phenomena via gain modulation.
The illusory simplicity of the feedforward pass — evidence for dynamical nature of stimulus encoding in neural networks. Demonstrates that feedforward computation involves complex temporal dynamics rather than static transformations. Applicable to neural…
Dynamical systems methodology for flexible phase-locking in cortical oscillators. Multi-timescale inhibitory currents enable entrainment to rhythms slower than intrinsic frequency via delayed Hopf bifurcation. Activation: phase-locking, cortical oscillators,…
Geometric State-Space Neural Network for brain dynamics modeling. Combines state-space models with geometric constraints on brain connectivity to capture latent neural state evolution. Applicable to fMRI/EEG dynamics modeling, functional neuroimaging…
Robust volatility updates for Hierarchical Gaussian Filtering (HGF). Improves stability and convergence of uncertainty estimation in perceptual inference. Activation: hierarchical gaussian filter, volatility update, perceptual inference, active inference,…