@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.002_pinet.yaml

@@ -0,0 +1,189 @@
+schema_version: 1
+id: aisb.t3.002_pinet
+name: 'PINet: Optimizing hard-constrained neural networks with orthogonal projection
+  layers'
+version: 0.1.0
+one_line: 'Train and benchmark feasible-by-design neural networks that enforce convex
+  constraints via Douglas–Rachford projection layers in JAX, measuring inference speed,
+  constraint violation, and relative suboptimality on parametric QP and toy-MPC problems.
+
+  '
+task_description: 'This benchmark evaluates ΠNet (PINet), a neural-network output
+  layer that orthogonally projects onto convex constraint sets using operator splitting
+  (Douglas–Rachford) in the forward pass and the implicit function theorem for backpropagation.
+  The packaged task trains backbone + projection models on parametric constrained
+  optimization problems (QP benchmarks at two scales: d=100 and d=1000, plus a toy
+  MPC problem), then measures single- and batch-inference time, constraint violation
+  (CV), and relative suboptimality (RS) against optimal solutions from high-accuracy
+  solvers. The small QP dataset is bundled; the large QP dataset must be obtained
+  from the paper''s OpenReview supplementary material. All training and evaluation
+  code is self-contained in the snapshot and runs via `src/benchmarks/QP/run_QP.py`
+  and `src/benchmarks/toy_MPC/run_toy_MPC.py`. A Docker Compose setup is provided
+  for reproducible CPU and GPU execution.
+
+  '
+capability_tags:
+- research_code_optimization
+- constrained_optimization
+- jax
+- scientific_ml
+- systems_efficiency
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+cost_band: low
+time_band: 1-2h
+difficulty: medium
+data_access: public
+primary_outputs:
+- convex_small_batch_inference_time_s
+- nonconvex_small_batch_inference_time_s
+- relative_suboptimality
+- constraint_violation
+- benchmark_report
+launch_profiles:
+- id: quick_check
+  label: Quick Check
+  description: 'Run the smallest QP benchmark script (small dataset, single seed)
+    to verify projection layers compile, train, and produce feasibility metrics. Expect
+    ~10 min on CPU.
+
+    '
+- id: full_benchmark
+  label: Full Benchmark
+  description: 'Run the full QP benchmark suite via run_QP_batch.sh across all configs,
+    seeds, and projection methods (PINet, JAXopt, cvxpylayers). Produces comparative
+    RS, CV, and timing results matching the paper. Expect ~1-2 h on GPU, longer on
+    CPU.
+
+    '
+- id: toy_mpc
+  label: Toy MPC
+  description: 'Generate the toy MPC dataset and run end-to-end training + evaluation
+    with run_toy_MPC.py. Produces training logs, validation metrics, and saved parameters.
+
+    '
+dataset_download:
+  primary_method: bundled
+  sources:
+  - kind: bundled
+    url: https://deepscientist.cc/AISB/002_pinet
+    access: public
+    note: 'Small QP benchmark data is included in the snapshot archive.
+
+      '
+  - kind: external
+    url: https://openreview.net
+    access: public
+    note: 'Large QP dataset (d=1000, n_eq=500, n_ineq=500) must be obtained from the
+      paper''s OpenReview supplementary material. Not bundled.
+
+      '
+  notes:
+  - Small dataset is bundled and sufficient for the quick_check and full_benchmark
+    profiles at small scale.
+  - Toy MPC dataset can be generated locally via `python -m src.benchmarks.toy_MPC.generate_toy_MPC`.
+  - Total disk footprint including code and small data is under 1 GB.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No API keys, tokens, or external accounts are required.
+  - Optional wandb integration exists in run_QP.py but is not mandatory.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 40
+    gpu_count: 1
+    gpu_vram_gb: 16
+environment:
+  python: '3.10'
+  cuda: '12'
+  pytorch: 2.6.0
+  key_packages:
+  - jax==0.6.2
+  - optax==0.2.4
+  - cvxpy==1.6.6
+  - flax
+  - pinet-hcnn
+  notes:
+  - CPU-only execution is fully supported for the minimum route; install via `pip
+    install pinet-hcnn`.
+  - GPU mode requires NVIDIA CUDA 12; install via `pip install "pinet-hcnn[cuda12]"`.
+  - Docker Compose configs provided for both CPU (`pinet-cpu`) and GPU (`pinet-gpu`)
+    reproducibility.
+  - NVIDIA Container Toolkit required on host for GPU Docker route.
+  - PyTorch is used only for data loading (torch DataLoader); JAX is the primary compute
+    framework.
+risk_flags:
+- large_dataset_external
+- optional_wandb
+risk_notes:
+- 'The large QP benchmark dataset (d=1000) is not bundled and must be downloaded from
+  OpenReview supplementary material. Without it, only the small-scale QP experiments
+  can be reproduced.
+
+  '
+- 'wandb logging is referenced in run_QP.py but is optional; runs complete without
+  it.
+
+  '
+- 'GPU CUDA wheel availability is Linux-only (x86_64 and aarch64). Windows and macOS
+  are CPU-only.
+
+  '
+recommended_when: 'Use this benchmark when you need a constrained-learning task exercising
+  implicit-layer projection methods, when you want to compare feasible-by-design approaches
+  on parametric QPs, or when you want a reproducible JAX benchmark that can run on
+  CPU-only hardware in under two hours.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you need large-scale vision, language,
+  or multimodal training workloads, or if you require non-convex constraint handling
+  beyond the objective function. Also not suitable if you cannot install JAX (e.g.,
+  non-Linux GPU environments).
+
+  '
+paper:
+  title: 'PINet: Optimizing hard-constrained neural networks with orthogonal projection
+    layers'
+  authors:
+  - Panagiotis D. Grontas
+  - Antonio Terpin
+  - Efe C. Balta
+  - Raffaello D'Andrea
+  - John Lygeros
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2508.10480
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.002_pinet.zip
+  archive_type: zip
+  local_dir_name: paper-2-PINet
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.002_pinet.zip
+  sha256: eb3944c02853116c2cf70f2213a7809eee8f0714143e5dac4f7e35876d187535
+  size_bytes: 180464
+commercial:
+  annual_fee: null
+display:
+  palette_seed: graphite-lime-precision
+  art_style: technical-minimal
+  accent_priority: medium
+image_path: ../image/002_aisb.t3.002_pinet.jpg

package/AISB/catalog/aisb.t3.002_pinet.zh.yaml

@@ -0,0 +1,130 @@
+schema_version: 1
+id: aisb.t3.002_pinet
+name: 'PINet：通过正交投影层优化硬约束神经网络'
+version: 0.1.0
+one_line: '在JAX中训练并基准测试通过Douglas–Rachford投影层强制凸约束的可行设计神经网络，测量推理速度、约束违反和相对次优性在参数化QP和玩具MPC问题上的表现。'
+task_description: '该基准测试评估ΠNet (PINet)，这是一种神经网络输出层，通过算子分裂（Douglas–Rachford）在前向传播中对凸约束集进行正交投影，并使用隐函数定理进行反向传播。打包的任务在参数化约束优化问题（QP基准测试两个规模：d=100和d=1000，以及一个玩具MPC问题）上训练骨干网络+投影模型，然后测量单次和批量推理时间、约束违反（CV）以及相对于高精度求解器最优解的相对次优性（RS）。小规模QP数据集已打包；大规模QP数据集需从论文的OpenReview补充材料获取。所有训练和评估代码均包含在快照中，通过`src/benchmarks/QP/run_QP.py`和`src/benchmarks/toy_MPC/run_toy_MPC.py`运行。提供了Docker Compose设置以确保CPU和GPU执行的可重现性。'
+capability_tags:
+- research_code_optimization
+- constrained_optimization
+- jax
+- scientific_ml
+- systems_efficiency
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+cost_band: low
+time_band: 1-2h
+difficulty: medium
+data_access: public
+primary_outputs:
+- convex_small_batch_inference_time_s
+- nonconvex_small_batch_inference_time_s
+- relative_suboptimality
+- constraint_violation
+- benchmark_report
+launch_profiles:
+- id: quick_check
+  label: 快速检查
+  description: '运行最小的QP基准测试脚本（小数据集、单次随机种子）以验证投影层编译、训练并产生可行性指标。CPU上预计约10分钟。'
+- id: full_benchmark
+  label: 完整基准测试
+  description: '通过run_QP_batch.sh运行完整QP基准测试套件，覆盖所有配置、种子和投影方法（PINet、JAXopt、cvxpylayers）。生成与论文匹配的RS、CV和计时对比结果。GPU上预计约1-2小时，CPU上更长。'
+- id: toy_mpc
+  label: 玩具MPC
+  description: '生成玩具MPC数据集并使用run_toy_MPC.py运行端到端训练+评估。生成训练日志、验证指标和保存的参数。'
+dataset_download:
+  primary_method: bundled
+  sources:
+  - kind: bundled
+    url: https://deepscientist.cc/AISB/002_pinet
+    access: public
+    note: '小规模QP基准测试数据已包含在快照压缩包中。'
+  - kind: external
+    url: https://openreview.net
+    access: public
+    note: '大规模QP数据集（d=1000、n_eq=500、n_ineq=500）需从论文的OpenReview补充材料获取。未打包。'
+  notes:
+  - 小数据集已打包，足以支持小规模的quick_check和full_benchmark配置。
+  - 玩具MPC数据集可通过`python -m src.benchmarks.toy_MPC.generate_toy_MPC`在本地生成。
+  - 包含代码和小数据的总磁盘占用小于1 GB。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 无需API密钥、令牌或外部账户。
+  - run_QP.py中存在可选的wandb集成，但非必需。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 40
+    gpu_count: 1
+    gpu_vram_gb: 16
+environment:
+  python: '3.10'
+  cuda: '12'
+  pytorch: 2.6.0
+  key_packages:
+  - jax==0.6.2
+  - optax==0.2.4
+  - cvxpy==1.6.6
+  - flax
+  - pinet-hcnn
+  notes:
+  - CPU仅执行模式已完全支持最小配置；通过`pip install pinet-hcnn`安装。
+  - GPU模式需要NVIDIA CUDA 12；通过`pip install "pinet-hcnn[cuda12]"`安装。
+  - 提供了CPU（pinet-cpu）和GPU（pinet-gpu）两种Docker Compose配置以确保可重现性。
+  - GPU Docker方式需要在主机上安装NVIDIA Container Toolkit。
+  - PyTorch仅用于数据加载（torch DataLoader）；JAX是主要计算框架。
+risk_flags:
+- large_dataset_external
+- optional_wandb
+risk_notes:
+- '大规模QP基准测试数据集（d=1000）未打包，需从OpenReview补充材料下载。没有该数据集，只能复现小规模QP实验。'
+- 'run_QP.py中引用了wandb日志记录但为可选；即使没有wandb运行也能完成。'
+- 'GPU CUDA wheel仅支持Linux（x86_64和aarch64）。Windows和macOS仅支持CPU。'
+recommended_when: '当你需要一个涉及隐式层投影方法的约束学习任务时、当你想在参数化QP上比较可行设计方法时、或者当你需要一个可在不到两小时内仅用CPU硬件运行的JAX基准测试时，使用此基准测试。'
+not_recommended_when: '如果需要大规模视觉、语言或多模态训练工作负载，或者需要处理目标函数以外的非凸约束，请勿使用此基准测试。同样不适用于无法安装JAX的环境（如非Linux GPU环境）。'
+paper:
+  title: 'PINet: Optimizing hard-constrained neural networks with orthogonal projection
+    layers'
+  authors:
+  - Panagiotis D. Grontas
+  - Antonio Terpin
+  - Efe C. Balta
+  - Raffaello D'Andrea
+  - John Lygeros
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2508.10480
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.002_pinet.zip
+  archive_type: zip
+  local_dir_name: paper-2-PINet
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.002_pinet.zip
+  sha256: eb3944c02853116c2cf70f2213a7809eee8f0714143e5dac4f7e35876d187535
+  size_bytes: 180464
+commercial:
+  annual_fee: null
+display:
+  palette_seed: graphite-lime-precision
+  art_style: technical-minimal
+  accent_priority: medium
+image_path: ../image/002_aisb.t3.002_pinet.jpg

package/AISB/catalog/aisb.t3.004_decentralattn.yaml

@@ -0,0 +1,184 @@
+schema_version: 1
+id: aisb.t3.004_decentralattn
+name: 'TeCh: Decentralized Attention Fails Centralized Signals'
+version: 0.1.0
+one_line: Train and evaluate the TeCh (CoTAR-based) medical time-series classification
+  framework across five EEG/ECG benchmark datasets under a subject-independent protocol.
+task_description: 'This benchmark evaluates the TeCh framework—a Transformer variant
+  that replaces standard quadratic self-attention with the linear-complexity CoTAR
+  (Core Token Aggregation-Redistribution) module—on medical time-series (MedTS) classification.
+  The task is to train and tune TeCh on up to five preprocessed clinical datasets
+  (3 EEG: APAVA, ADFTD, TDBrain; 2 ECG: PTB, PTB-XL) using the subject-independent
+  evaluation protocol, where train/val/test sets are split by patient so the model
+  must generalize to unseen subjects. Each dataset has a dedicated shell script under
+  ./scripts/ that launches run.py → exp/exp_classification.py with dataset-specific
+  hyperparameters (batch size, learning rate, number of temporal/channel encoder layers,
+  augmentation mix, epoch budget). The model outputs classification logits; evaluation
+  computes accuracy, precision, recall, F1, AUROC, and AUPRC. Datasets must be downloaded
+  externally from the Medformer repository and placed under ./dataset/. The TDBrain
+  dataset additionally requires institutional permission. The goal is to improve classification
+  quality (primarily accuracy and F1) by modifying architecture, training loop, or
+  hyperparameters without breaking the multi-dataset evaluation route.
+
+  '
+capability_tags:
+- research_code_optimization
+- medical_time_series
+- transformers
+- classification
+- health_ai
+- eeg
+- ecg
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: restricted
+primary_outputs:
+- accuracy
+- precision
+- recall
+- f1
+- auroc
+- auprc
+- model_checkpoint
+- training_logs
+launch_profiles:
+- id: quick_check
+  label: Quick Check (Single Dataset)
+  description: 'Run a single dataset script (e.g., bash ./scripts/APAVA.sh) with default
+    hyperparameters to verify the training loop, data loading, and metric computation
+    work end-to-end. Runs 5 iterations × 40 epochs on APAVA; expect 1–3 hours on a
+    single GPU.
+
+    '
+- id: full_eval
+  label: Full Benchmark Suite
+  description: 'Run all dataset scripts (APAVA, ADFTD, FLAAP, PTB, PTB-XL, and optionally
+    TDBrain) sequentially or in parallel. Each script runs 5 random-seed iterations.
+    Expect 6–24 hours total on a single GPU depending on dataset sizes and epoch budgets.
+
+    '
+- id: ablation
+  label: Ablation / Architecture Modification
+  description: 'Modify the TeCh architecture (e.g., number of temporal vs. channel
+    encoder layers, CoTAR core-token dimension, augmentation strategies) and re-run
+    the full suite to measure impact on all six classification metrics.
+
+    '
+dataset_download:
+  primary_method: external
+  sources:
+  - kind: github
+    url: https://github.com/DL4mHealth/Medformer
+    access: public
+    note: 'All preprocessed MedTS datasets (APAVA, ADFTD, FLAAP, PTB, PTB-XL, UCI-HAR)
+      are hosted or linked from the Medformer repository. Download and place under
+      ./dataset/.
+
+      '
+  - kind: restricted
+    url: https://github.com/DL4mHealth/Medformer
+    access: permission_required
+    note: 'The TDBrain dataset requires institutional permission before download.
+      See Medformer repo for access instructions.
+
+      '
+  notes:
+  - Shell scripts hardcode data paths like /data/gqyu/dataset/med/APAVA/; update --root_path
+    in each script to match your local layout under ./dataset/.
+  - Total disk footprint for all datasets is estimated at 5–30 GB depending on which
+    datasets are included.
+credential_requirements:
+  mode: conditional
+  items:
+  - TDBrain dataset requires institutional data access permission
+  notes:
+  - Five of the six datasets are publicly available through the Medformer repository.
+  - TDBrain requires a separate access request; the benchmark is fully functional
+    without it.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 120
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: 2.4.1
+  flash_attn: null
+  key_packages:
+  - scikit-learn
+  - psutil
+  - numpy
+  notes:
+  - See requirements.txt for the full dependency set.
+  - Code uses psutil for CPU-affinity pinning; ensure it is installed.
+  - CUDA version not explicitly specified; any version compatible with PyTorch 2.4.1
+    should work.
+risk_flags:
+- external_dataset_dependency
+- hardcoded_paths
+- restricted_data_subset
+risk_notes:
+- Datasets are not bundled in the snapshot; they must be downloaded from the Medformer
+  repository and placed under ./dataset/.
+- Shell scripts contain hardcoded absolute paths (e.g., /data/gqyu/dataset/med/APAVA/)
+  that must be edited before running.
+- The TDBrain dataset requires institutional permission; omitting it still leaves
+  five runnable benchmarks.
+- Each shell script runs 5 random-seed iterations (--itr 5), which multiplies wall-clock
+  time; reduce for quick validation.
+- num_workers defaults to 16 in run.py; reduce if fewer CPU cores are available.
+recommended_when: 'Use this benchmark when you want to test architecture or training-loop
+  modifications for medical time-series classification on clinically relevant EEG/ECG
+  data under a rigorous subject-independent evaluation protocol, with six standard
+  classification metrics computed automatically.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you cannot obtain the external
+  datasets from the Medformer repository, if you need a purely self-contained data
+  route with no external downloads, or if your research focus is time-series forecasting
+  rather than classification.
+
+  '
+paper:
+  title: 'Decentralized Attention Fails Centralized Signals: Rethinking Transformers
+    for Medical Time Series'
+  venue: ICLR 2026 (Oral)
+  year: 2026
+  url: https://openreview.net/forum?id=oZJFY2BQt2
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.004_decentralattn.zip
+  archive_type: zip
+  local_dir_name: paper-4-DecentralAttn
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.004_decentralattn.zip
+  sha256: bf0dd8489ed4694dc4c1e494d25e9f564d99813e669ded0cef9a85cb3a2ae36c
+  size_bytes: 71932
+commercial:
+  annual_fee: null
+display:
+  palette_seed: crimson-teal-clinic
+  art_style: signal-lab
+  accent_priority: high
+image_path: ../image/004_aisb.t3.004_decentralattn.jpg

package/AISB/catalog/aisb.t3.004_decentralattn.zh.yaml

@@ -0,0 +1,153 @@
+schema_version: 1
+id: aisb.t3.004_decentralattn
+name: 'TeCh：去中心化注意力机制无法捕获集中化信号'
+version: 0.1.0
+one_line: 在五个EEG/ECG基准数据集上，基于主题无关协议训练和评估TeCh（基于CoTAR）医学时间序列分类框架。
+task_description: '本基准测试评估TeCh框架——一种用线性复杂度的CoTAR（核心令牌聚合-重分配）模块替代标准二次自注意力的Transformer变体——在医学时间序列（MedTS）分类任务上的表现。任务是在最多五个预处理的临床数据集（3个EEG：APAVA、ADFTD、TDBrain；2个ECG：PTB、PTB-XL）上训练和调优TeCh，使用主题无关评估协议，其中训练集/验证集/测试集按患者划分，因此模型必须能够泛化到未见过的受试者。每个数据集在./scripts/下有专门的shell脚本，用于启动run.py → exp/exp_classification.py，并带有数据集特定的超参数（批大小、学习率、时间/通道编码器层数、增强混合、epoch预算）。模型输出分类logits；评估计算准确率、精确率、召回率、F1、AUROC和AUPRC。数据集必须从Medformer仓库外部下载并放置在./dataset/下。TDBrain数据集还需要机构许可。目标是改进分类质量（主要是准确率和F1），通过修改架构、训练循环或超参数，同时不破坏多数据集评估流程。
+
+  '
+capability_tags:
+- research_code_optimization
+- medical_time_series
+- transformers
+- classification
+- health_ai
+- eeg
+- ecg
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: restricted
+primary_outputs:
+- accuracy
+- precision
+- recall
+- f1
+- auroc
+- auprc
+- model_checkpoint
+- training_logs
+launch_profiles:
+- id: quick_check
+  label: 快速检查（单数据集）
+  description: '运行单个数据集脚本（如 bash ./scripts/APAVA.sh），使用默认超参数来验证训练循环、数据加载和指标计算是否端到端正常工作。在APAVA上运行5次迭代×40个epoch；在单个GPU上预计需要1-3小时。
+
+    '
+- id: full_eval
+  label: 完整基准测试套件
+  description: '按顺序或并行运行所有数据集脚本（APAVA、ADFTD、FLAAP、PTB、PTB-XL，以及可选的TDBrain）。每个脚本运行5个随机种子迭代。在单个GPU上根据数据集大小和epoch预算，预计总共需要6-24小时。
+
+    '
+- id: ablation
+  label: 消融实验/架构修改
+  description: '修改TeCh架构（如时间编码器与通道编码器的层数、CoTAR核心令牌维度、增强策略），并重新运行完整套件以测量对所有六个分类指标的影响。
+
+    '
+dataset_download:
+  primary_method: external
+  sources:
+  - kind: github
+    url: https://github.com/DL4mHealth/Medformer
+    access: public
+    note: '所有预处理的MedTS数据集（APAVA、ADFTD、FLAAP、PTB、PTB-XL、UCI-HAR）都托管在Medformer仓库中或从其中链接。下载并放置在./dataset/下。
+
+      '
+  - kind: restricted
+    url: https://github.com/DL4mHealth/Medformer
+    access: permission_required
+    note: 'TDBrain数据集需要机构许可才能下载。详见Medformer仓库的访问说明。
+
+      '
+  notes:
+  - Shell脚本硬编码了数据路径如/data/gqyu/dataset/med/APAVA/；请更新每个脚本中的--root_path以匹配您在./dataset/下的本地布局。
+  - 所有数据集的总磁盘占用估计为5-30 GB，取决于包含哪些数据集。
+credential_requirements:
+  mode: conditional
+  items:
+  - TDBrain数据集需要机构数据访问许可
+  notes:
+  - 六个数据集中有五个通过Medformer仓库公开可用。
+  - TDBrain需要单独的访问请求；即使不使用它，基准测试仍可完全运行。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 120
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: 2.4.1
+  flash_attn: null
+  key_packages:
+  - scikit-learn
+  - psutil
+  - numpy
+  notes:
+  - 完整依赖项列表请参见requirements.txt。
+  - 代码使用psutil进行CPU亲和性固定；请确保已安装。
+  - 未明确指定CUDA版本；任何与PyTorch 2.4.1兼容的版本都应该可以工作。
+risk_flags:
+- external_dataset_dependency
+- hardcoded_paths
+- restricted_data_subset
+risk_notes:
+- 数据集未打包在快照中；必须从Medformer仓库下载并放置在./dataset/下。
+- Shell脚本包含硬编码的绝对路径（如/data/gqyu/dataset/med/APAVA/），运行前必须编辑。
+- TDBrain数据集需要机构许可；省略它仍有五个可运行的基准测试。
+- 每个shell脚本运行5个随机种子迭代（--itr 5），这会使实际运行时间成倍增加；可减少迭代次数以快速验证。
+- num_workers在run.py中默认为16；如果CPU核心数较少，请减少此值。
+recommended_when: '当您想在严格的的主题无关评估协议下，在临床相关的EEG/ECG数据上测试医学时间序列分类的架构或训练循环修改，并自动计算六个标准分类指标时，请使用此基准测试。
+
+  '
+not_recommended_when: '如果无法从Medformer仓库获取外部数据集，如果您需要一个完全自包含且无外部下载的数据流程，或者如果您的研究重点是时间序列预测而非分类，则不要使用此基准测试。
+
+  '
+paper:
+  title: 'Decentralized Attention Fails Centralized Signals: Rethinking Transformers
+    for Medical Time Series'
+  venue: ICLR 2026 (Oral)
+  year: 2026
+  url: https://openreview.net/forum?id=oZJFY2BQt2
+  abstract_summary: '本论文重新审视了Transformer架构在医学时间序列分类中的应用，发现标准去中心化注意力机制无法有效捕获临床信号中的全局模式。我们提出了TeCh框架，通过核心令牌聚合-重分配（CoTAR）模块实现线性复杂度的注意力计算，在五个医学基准数据集上实现了最先进的性能。
+
+    '
+  abstract: 'We re-examine the Transformer architecture for medical time-series classification and find that standard decentralized attention mechanisms fail to effectively capture global patterns in clinical signals. We propose the TeCh framework with the Core Token Aggregation-Redistribution (CoTAR) module that achieves linear-complexity attention computation and state-of-the-art performance across five medical benchmark datasets.
+
+    '
+  notes:
+  - 本论文被ICLR 2026接收为Oral报告。
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.004_decentralattn.zip
+  archive_type: zip
+  local_dir_name: paper-4-DecentralAttn
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.004_decentralattn.zip
+  sha256: bf0dd8489ed4694dc4c1e494d25e9f564d99813e669ded0cef9a85cb3a2ae36c
+  size_bytes: 71932
+commercial:
+  annual_fee: null
+display:
+  palette_seed: crimson-teal-clinic
+  art_style: signal-lab
+  accent_priority: high
+image_path: ../image/004_aisb.t3.004_decentralattn.jpg