Menu
Skills dans ce dépôt
SkillsMP a collecté 4 114 skills depuis hiyenwong/ai_collection. Ouvrez un skill pour examiner sa source et ses détails.
hiyenwong/ai_collectionAffichage de 40 skills collectés sur 4 114.
Preisach Attention Layer (PAL) — a novel sequence modeling architecture that replaces softmax attention with the classical Preisach hysteresis operator from mathematical physics. Uses binary relay operators with learned thresholds and a stack of local extrema…
Langue du texte source : anglais
Universal prompt optimization skill that applies systematic techniques to improve prompt quality, clarity, and effectiveness for any AI agent.
Langue du texte source : anglais
Parallel multi-circuit quantum feature fusion methodology for medical image classification. Use when: (1) building hybrid quantum-classical CNN architectures for biomedical image classification, (2) comparing quantum vs classical models with statistical rigor…
Langue du texte source : anglais
Quantum Physics-Informed Neural Networks (QPINN) for portfolio optimization PDEs — uses parameterized quantum circuits with tensor rank decomposition to solve financial PDEs with 80x fewer parameters and higher accuracy than classical PINNs. Use when: solving…
Langue du texte source : anglais
Quasi-equivariant metanetworks for weight-space learning. Use for designing neural architectures that operate on pretrained model parameters, implementing equivariant and quasi-equivariant transformations that respect architectural symmetries while…
Langue du texte source : anglais
QUBO-based Byzantine-resilient federated learning methodology. Reformulates client selection in federated learning as Quadratic Unconstrained Binary Optimization solved by quantum annealers. Jointly optimizes over all client subsets to find mutually closest…
Langue du texte source : anglais
Research methodology from paper 'RAD-2: Scaling Reinforcement Learning in a Generator-Discriminator Framework'. arXiv:2604.15308v1. Covers key techniques and approaches for neuroscience research. Activation: rad, 2, scaling, cs.CV
Langue du texte source : anglais
Random graph-based datacenter network design (RNG) for cost-optimized, fault-tolerant distributed systems. Covers distributed routing protocols exploiting random graph properties, passive optical cabling shuffles, and production deployment patterns. Use when:…
Langue du texte source : anglais
RecursiveMAS framework for scaling multi-agent collaboration through recursive latent-space computation. Enables heterogeneous agents to form collaboration loops via RecursiveLink module with inner-outer loop learning. Activation triggers: recursive…
Langue du texte source : anglais
Relaxation-Informed Training (RIT) for neural network surrogate models enabling exact MILP embedding with reduced binary variables. Activation triggers: surrogate optimization, MILP embedding, neural network relaxation, ReLU pruning, optimization surrogate
Langue du texte source : anglais
Resolvent-RNN (R-RNN) methodology for constraining multi-hop temporal pathways in recurrent neural networks to achieve temporal sparsity alignment. Use when: (1) analyzing or designing RNN architectures with multi-hop temporal dependencies, (2) studying…
Langue du texte source : anglais
Using Reinforcement Fine-Tuning (RFT/GRPO) to overcome catastrophic forgetting in visual continual learning. Activation triggers: reinforcement fine-tuning continual learning, GRPO visual CL, RL visual continual learning, catastrophic forgetting visual,…
Langue du texte source : anglais
Methodology for analyzing how recurrent neural networks with neuron-specific adaptive time constants switch between multiple frequency band rhythms. Covers rhythm-switching mechanisms, time constant-frequency relationships, and degeneracy of learned…
Langue du texte source : anglais
RiM (Reasoning in Memory) methodology for unlocking working memory capacity in LLMs via fixed memory blocks, enabling compute-efficient latent reasoning without autoregressive thought generation.
Langue du texte source : anglais
RNN权重初始化、解的多样性与性能退化分析框架。研究不同初始化如何收敛到不同动力学解，分析网络规模、时间间隔、连接损伤对性能的优雅退化影响。适用于计算神经科学、RNN模型分析、脑皮层建模。触发词：RNN初始化、解多样性、性能退化、网络鲁棒性、优雅退化、weight initialization、degradation analysis、RNN dynamics、graceful degradation。
Langue du texte source : Plusieurs langues
Routing analysis and intervention for Multimodal Mixture-of-Experts models. Use when: (1) Debugging vision-language reasoning failures, (2) Analyzing expert routing in MoE architectures, (3) Improving multimodal MoE performance, (4) Understanding cross-modal…
Langue du texte source : anglais
Sample complexity bounds and algorithms for learning bosonic Gaussian quantum states. Use when: (1) analyzing sample requirements for quantum state tomography, (2) designing efficient measurement strategies for Gaussian states, (3) computing sample complexity…
Langue du texte source : anglais
Search-E1 methodology from arXiv:2605.22511 (May 2026). Self-evolution for search-augmented reasoning agents via vanilla GRPO + Offline Self-Distillation (OFSD) with token-level forward KL objective. No external supervision needed. Use when: training…
Langue du texte source : anglais
Biologically-inspired selective forgetting framework for LLM agent memory management. Combines hippocampal indexing theory and Ebbinghaus forgetting curve for efficient, secure, and quality-preserving memory pruning. Use for: agent memory optimization,…
Langue du texte source : anglais
Self-organising transformer architectures that determine their own structure during training. DDCL-INCRT pattern with hierarchical prototype structure. Use when designing adaptive neural architectures, self-organising networks, or architectures that evolve…
Langue du texte source : anglais
Formalizes how attractor networks emerge from the free energy principle applied to universal partitioning of random dynamical systems. Results in self-orthogonalizing attractor representations, biologically plausible multi-level Bayesian active inference. Use…
Langue du texte source : anglais
Self-Policy Distillation (SPD) methodology from arXiv:2605.22675 (May 2026). Capability-selective self-distillation by extracting low-rank subspace from gradients on correctness-defining tokens, projecting KV activations during self-generation, and…
Langue du texte source : anglais
Biologically plausible local self-supervised learning rules that learn hidden hierarchical data structure as efficiently as supervised backprop. Demonstrates that Direct Feedback Alignment (DFA) methods fail on hierarchical tasks due to input-specific…
Langue du texte source : anglais
Layerwise self-supervised local learning rules for deep networks on the Random Hierarchy Model (RHM). Direct feedback approximations fail due to input-specific masking nonlinearity; layerwise contrastive/non-contrastive self-supervised rules succeed and match…
Langue du texte source : anglais
Sensation Modulating Network (SMN) architecture for embodied cognition - haltability-based framework reconciling cognitivism and 4E approaches through opponent dynamics. Use when studying embodied cognitive architecture, intentional directedness without…
Langue du texte source : anglais
Sequential chaotic oscillations (SCOs) in excitatory-inhibitory threshold-linear networks - dynamical mechanism for sequential metastability in brain dynamics. Activation: sequential metastability, chaotic itinerancy, E-I oscillation, SCO, threshold-linear…
Langue du texte source : anglais
Sleep-like consolidation mechanism for LLMs that converts recent context into persistent fast weights before clearing KV cache, enabling long-horizon reasoning with preserved inference latency.
Langue du texte source : anglais
梯度突触可塑性稀疏实现方法论。实现真正稀疏且通用的梯度突触可塑性规则，保持在线学习能力。适用于脉冲神经网络、在线学习、突触可塑性研究。触发词：突触可塑性、梯度下降、稀疏实现、在线学习、sparse plasticity、gradient-based、online learning。
Langue du texte source : chinois
Workflow and architecture patterns for building robust AI agents using Specs, Contracts, and Repository patterns.
Langue du texte source : anglais
特异性感知联邦图学习框架(SFGL)。在保护数据隐私的前提下进行多站点fMRI协作训练，通过共享分支和个性化分支平衡知识共享与站点特异性保持。适用于多中心脑疾病识别、联邦学习、rs-fMRI分析。触发词：联邦学习、图神经网络、fMRI分析、多站点、数据隐私、federated learning、GNN、multi-site、brain disorder identification。
Langue du texte source : chinois
Optimize LLM inference using speculative decoding with KV cache compression. Covers adaptive gamma selection, compression-aware token verification, and KV cache management for reduced latency. Use when: (1) Optimizing LLM serving throughput, (2) Implementing…
Langue du texte source : anglais
Spike-driven Large Language Model (SDLLM) methodology. Eliminates dense matrix multiplications in LLMs through sparse addition operations using gamma-SQP two-step spike encoding. Reduces energy consumption by 7x while improving accuracy by 4.2% over previous…
Langue du texte source : anglais
Spike-driven Large Language Model - Spike-based computation for large language models. Activation triggers: spike, driven, large, neuroscience, SNN.
Langue du texte source : anglais
Spike-NVPT: Noise-robust visual prompt tuning using bio-inspired temporal filtering and spike-based discretization. Parameter-efficient adaptation for pre-trained vision models with enhanced robustness to input perturbations. Keywords: Spike-NVPT, visual…
Langue du texte source : anglais
SPOQ (Specialist Orchestrated Queuing) - Multi-agent software engineering methodology with wave-based topological dispatch, dual validation gates, and Human-as-an-Agent integration for automated SE tasks. Activation: multi-agent orchestration, software…
Langue du texte source : anglais
Staged VLM post-training methodology - decomposing VLM capabilities into visual perception, visual reasoning, and textual reasoning stages with specialized data and RL-based perception learning
Langue du texte source : anglais
Stateful streaming transformer inference methodology with persistent KV cache, Flash Queries prefetching, and multi-tenant continuous-batching for efficient streaming workloads.
Langue du texte source : anglais
Methodology for improving LLM reasoning by making search tree structures explicit in reasoning traces. LinTree (arXiv:2605.31492) shows that adding parent pointers to linearized search traces significantly outperforms implicit reasoning and LLM-heuristic…
Langue du texte source : anglais
Structured-Sparse Attention methodology from arXiv:2605.22476 (May 2026). Blockwise resolvent-style attention operator achieving subquadratic sequence complexity O(n^(4/3)) for entity tracking by exploiting localized attention structure. Use when: efficient…
Langue du texte source : anglais
Superintelligent Retrieval Agent methodology for building retrieval-augmented systems that actively reason about information needs beyond black-box query issuance.
Langue du texte source : anglais