@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.054_timebase.zh.yaml

@@ -0,0 +1,158 @@
+schema_version: 1
+id: aisb.t3.054_timebase
+name: 'TimeBase：极简主义在高效长期时间序列预测中的力量'
+version: 0.1.0
+one_line: >
+  训练并评估 TimeBase，这是一款超轻量级（约 0.39k 参数）的长期时间序列预测模型，
+  采用基提取和分段预测方法，在 17+ 个公开基准数据集上测量 MSE 和 MAE。
+task_description: |
+  本基准测试打包了 TimeBase，这是一款极轻量级的长期时间序列预测（LTSF）模型，
+  利用周期时间序列的近似低秩结构。模型将输入分割为长度为 P 的非重叠窗口，
+  通过单个线性层提取 R 个正交基分量，再通过第二个线性层预测未来片段——总计约 390 个参数。
+  智能体需要：
+  (1) 从 Google Drive 下载 17 个标准规模基准数据集（ETTh1/h2、ETTm1/m2、Weather、Electricity、
+  Traffic、Solar-Energy、Wind、Exchange-Rate、METR-LA、ZafNoo、CzeLan、AQShunyi、AQWan、PM2.5、Temp）
+  并放置在 ./dataset/ 目录中；
+  (2) 通过 `sh ./scripts/$DATA_NAME.sh` 运行每个数据集的训练脚本，
+  该脚本调用 `run_longExp.py` → `exp/exp_main.py`；
+  (3) 收集各预测 horizon（96、192、336、720）的 avg_mse 和 avg_mae。
+  超参数搜索范围包括 λ_orth ∈ [0.00–0.20] 和学习率 ∈ [0.01–0.5]。
+  可选的即插即用路线通过 `plug-and-play_for_patchtst/run_all.sh` 将 TimeBase 作为复杂度缩减器应用于 PatchTST。
+  指标由 `utils/metrics.py` 使用基于 numpy 的 MSE/MAE 计算。
+  所有评估均为自包含的，无需外部评估服务器。
+  此外，还可以从 LargeST 仓库获取 4 个大规模数据集（CA 4.52B、GLA 2.02B、GBA 1.24B、SD 0.38B 条记录）
+  用于扩展评估。
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- efficient_models
+- forecasting
+- scientific_ml
+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: medium
+time_band: 6-24h
+difficulty: medium
+data_access: public
+primary_outputs:
+- avg_mse
+- avg_mae
+- forecast_report
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: |
+    在单个标准规模数据集（如 ETTh1）上运行 TimeBase 训练和评估，
+    使用默认超参数。在单个 GPU 上不到 1 小时完成。
+- id: full_benchmark
+  label: 完整 17 数据集基准测试
+  description: |
+    依次运行全部 17 个标准规模数据集脚本，收集四个预测 horizon（96、192、336、720）
+    的 avg_mse 和 avg_mae。在单个 GPU 上预计需要 6–24 小时。
+- id: patchtst_plugin
+  label: PatchTST 即插即用
+  description: |
+    通过 plug-and-play_for_patchtst/run_all.sh 运行 PatchTST 的即插即用复杂度缩减器。
+    展示 77–93% 的 MAC 算力降低。
+- id: large_scale
+  label: 大规模数据集
+  description: |
+    将评估扩展到包含数十亿条记录的 4 个 LargeST 数据集（CA、GLA、GBA、SD）。
+    需要从 LargeST GitHub 仓库单独下载数据，并需要大量磁盘空间和计算时间。
+dataset_download:
+  primary_method: manual
+  sources:
+  - kind: google_drive
+    url: https://drive.google.com/file/d/1ypgCc6iQ2Z8IB_9CY3If_KMRNQKBsI3J/view?usp=sharing
+    access: public
+    note: |
+      17 个标准规模基准数据集（ETTh1/h2、ETTm1/m2、Weather、Electricity、Traffic、
+      Solar-Energy、Wind、Exchange-Rate、METR-LA、ZafNoo、CzeLan、AQShunyi、AQWan、
+      PM2.5、Temp）。解压至 ./dataset/ 目录。
+  - kind: github
+    url: https://github.com/liuxu77/LargeST
+    access: public
+    note: |
+      4 个大规模数据集（CA、GLA、GBA、SD），包含数十亿条记录。
+      按照 LargeST 预处理说明操作。仅在大规模数据集启动配置时需要。
+  notes:
+  - 标准规模数据集体积适中（解压后总计约几百 MB）。
+  - 大规模数据集需要数百 GB 磁盘空间，且为可选项。
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 1000
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: 3.8+
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - numpy
+  - torch
+  - matplotlib
+  notes:
+  - 该模型极其轻量级（约 390 个参数）；GPU 内存不是瓶颈。
+  - 推荐使用 1 TB 磁盘以容纳可选的大规模 LargeST 数据集。
+  - 完整的依赖项列表请参阅打包的 requirements 文件。
+  - 仓库中未固定 Python、CUDA 和 PyTorch 版本；任何支持 CUDA 的近期 PyTorch 版本均可工作。
+risk_flags:
+- manual_dataset_download
+- large_dataset_route
+risk_notes:
+- 数据集必须从 Google Drive 手动下载并解压至 ./dataset/ 目录，否则训练脚本将无法运行。
+- 4 个大规模 LargeST 数据集涉及数十亿条记录，需要单独的预处理步骤；
+  在磁盘和内存不足的情况下尝试会导致失败。
+- 未固定依赖版本；对于较旧或较新的 Python/PyTorch 组合，可能需要手动解决环境问题。
+recommended_when: |
+  当您需要一个以效率为重点的长期时间序列预测任务，且使用成熟的公开数据集时，
+  或者研究 LTSF 的极端模型压缩（子 1k 参数）时，请使用此基准测试。
+  非常适合需要一个轻量级、自包含的训练+评估循环以及明确的数值指标（MSE、MAE）的智能体。
+not_recommended_when: |
+  如果您需要多模态基准测试、基础模型规模的任务，
+  或以大型预训练模型为主的基准测试，请勿使用此基准测试。
+  如果您无法从 Google Drive 下载数据，则不适用。
+  没有数百 GB 磁盘空间，大规模配置不切实际。
+paper:
+  title: 'TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting'
+  venue: ICML 2025 Spotlight
+  year: 2025
+  url: https://proceedings.mlr.press/v267/huang25az.html
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.054_timebase.zip
+  archive_type: zip
+  local_dir_name: paper-54-TimeBase
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.054_timebase.zip
+  sha256: 0871383ee8a2d7dc7f00797e86c0392047d967f3824bf6e76052081150baadce
+  size_bytes: 346725
+commercial:
+  annual_fee: null
+display:
+  palette_seed: ivory-blue-minimal
+  art_style: minimal-forecast
+  accent_priority: medium
+image_path: ../image/054_aisb.t3.054_timebase.jpg

package/AISB/catalog/aisb.t3.055_csbrain.yaml

@@ -0,0 +1,238 @@
+id: aisb.t3.055_csbrain
+name: 'CSBrain: Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding'
+version: 0.1.0
+one_line: Cross-scale EEG foundation model with CST tokenization and SSA attention
+  for generalized brain decoding across 16 datasets.
+task_description: 'CSBrain is a cross-scale spatiotemporal brain foundation model
+  for EEG decoding, introduced at NeurIPS 2025. The model addresses limitations of
+  prior EEG foundation models that adopt scale-agnostic dense modeling by introducing
+  two key components: (i) Cross-scale Spatiotemporal Tokenization (CST), which aggregates
+  multi-scale EEG features within localized temporal windows and anatomical brain
+  regions into compact scale-aware token representations; and (ii) Structured Sparse
+  Attention (SSA), which captures long-range dependencies across temporal windows
+  and brain regions while minimizing spurious correlations common in noisy EEG signals.
+  CST and SSA are alternately stacked to progressively integrate cross-scale spatiotemporal
+  dependencies. The benchmark supports both pretraining from raw EEG signals and downstream
+  fine-tuning across 16 datasets spanning 11 representative EEG tasks including emotion
+  recognition, motor imagery, seizure detection, and sleep staging. Environment setup
+  delegates to CBraMod framework; pretrained weights are available via Google Drive.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 1d+
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- balanced_accuracy
+- cohens_kappa
+- weighted_f1
+- eeg_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run evaluation on a single downstream dataset using provided fine-tuned
+    weights. Loads pretrained CSBrain checkpoint and evaluates on one of 16 supported
+    datasets (e.g., BCIC-IV-2a, CHB-MIT, TUAB, SEED-V). Fastest path to metric collection.
+
+    '
+- id: pretrain_and_finetune
+  label: Pretrain + Finetune
+  description: 'Full pipeline: pretrain CSBrain from scratch using masked autoencoding
+    on large-scale EEG data, then fine-tune on downstream datasets. Supports 16 datasets
+    including BCIC-IV-2a, Siena, HMC, TUSL, and all CBraMod benchmarks. Default training
+    uses 40 epochs, batch size 128, learning rate 5e-4.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: CBraMod benchmarks
+    type: benchmark_package
+    notes: '16 downstream datasets supported including BCIC-IV-2a, BCIC-IV-2b, CHB-MIT,
+      TUAB, TUEV, SEED-V, SEED-VIG, ISRUC-Sleep, PhysioNet, FACED, Mumtaz2016, SHU,
+      Siena, HMC, TUSL, and stress.
+
+      '
+  - name: Pretraining dataset
+    type: external
+    notes: Follow CBraMod preprocessing pipeline for pretraining data preparation.
+  - name: Google Drive pretrained weights
+    type: external
+    url: https://drive.google.com/drive/folders/1-GsVVewRM0B93H08yts5m53yU2whxYvj
+    notes: 'CSBrain.pth contains pretrained foundation model weights; pth_downtasks/
+      contains fine-tuned weights for all 16 downstream datasets. Recommended for
+      evaluation without retraining.
+
+      '
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - Google Drive access may require authentication for weight downloads.
+  - CBraMod framework access required for full dataset pipeline.
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: Single GPU sufficient for evaluation only; pretraining requires significantly
+      more resources.
+  recommended:
+    cpu_cores: 64
+    ram_gb: 256
+    disk_gb: 500
+    gpu_count: 4
+    gpu_vram_gb: 80
+    notes: 'Multi-GPU setup recommended for pretraining (supports parallel=True in
+      pretrain_main.py). Fine-tuning on individual datasets can run on single A100
+      80GB; full multi-dataset evaluation benefits from 4x80GB configuration.
+
+      '
+environment:
+  python: 3.8+
+  cuda: 11.7+
+  pytorch: 2.0+
+  flash_attn: null
+  key_packages:
+  - torch
+  - numpy
+  - scikit-learn
+  - CBraMod
+  notes:
+  - Environment setup delegates entirely to CBraMod framework per bundled README instructions.
+  - Install CBraMod dependencies first, then CSBrain-specific requirements if any.
+  - Preprocessing scripts provided in `preprocessing/` folder for Siena, HMC, and
+    TUSL datasets.
+  - No standalone requirements.txt; dependency resolution handled through CBraMod.
+risk_flags:
+- large_pretraining_cost
+- multi_dataset_evaluation_overhead
+- google_drive_access
+risk_notes:
+- Pretraining CSBrain from scratch is computationally expensive; downloading pretrained
+  weights recommended for most evaluation scenarios.
+- Running full fine-tuning across all 16 datasets requires significant disk space
+  and time (estimated 2-4 days on recommended hardware).
+- Google Drive weights may require manual authentication or VPN access depending on
+  geographic location.
+- CBraMod dependency may introduce version conflicts with other benchmarks in the
+  same environment.
+recommended_when: 'Use this benchmark when you need a brain-signal foundation model
+  benchmark with extensive EEG transfer learning evaluation. Ideal for research on
+  cross-scale representations, structured sparse attention, or generalization across
+  heterogeneous EEG datasets. Appropriate for healthcare AI research, BCI development,
+  and scientific machine learning exploration.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you require a lightweight baseline,
+  have limited GPU resources (<24GB VRAM), or cannot host multiple EEG datasets and
+  checkpoints (500GB+ recommended). Not suitable for environments lacking CBraMod
+  framework support or where Google Drive access is restricted.
+
+  '
+paper:
+  title: 'CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding'
+  venue: NeurIPS 2025 Spotlight
+  year: 2025
+  authors:
+  - Yuchen Zhou
+  - Jiamin Wu
+  - Zichen Ren
+  - Zhouheng Yao
+  - Weiheng Lu
+  - Kunyu Peng
+  - Qihao Zheng
+  - Chunfeng Song
+  - Wanli Ouyang
+  - Chao Gou
+  url: https://arxiv.org/abs/2506.23075
+  github: https://github.com/yuchen2199/CSBrain
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.055_csbrain.zip
+  archive_type: zip
+  local_dir_name: paper-55-CSBrain
+  notes:
+  - Includes sh/ folder with fine-tuning scripts for all 16 datasets.
+  - Pretrained weights (CSBrain.pth) and downstream weights (pth_downtasks/) downloaded
+    separately from Google Drive.
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.055_csbrain.zip
+  sha256: 3d02943b88e62e967308b3cc9bae379cab2431139fcfef0f591fbc95fb206d4c
+  size_bytes: 123771
+display:
+  palette_seed: violet-cyan-neuro
+  art_style: brain-atlas
+  accent_priority: high
+image_path: ../image/055_aisb.t3.055_csbrain.jpg
+track_fit:
+- paper_track
+- benchmark_track
+capability_tags:
+- research_code_optimization
+- eeg_decoding
+- brain_foundation_models
+- scientific_ml
+- healthcare_ai
+aisb_direction: T3
+code_anchors:
+  evaluation:
+  - path: evaluate_bci2a_v2.py
+    description: Primary evaluation script for BCIC-IV-2a dataset
+    metrics:
+    - balanced_accuracy
+    - cohens_kappa
+    - weighted_f1
+  training:
+  - path: pretrain_main.py
+    description: Foundation model pretraining entry point
+    defaults:
+      epochs: 40
+      batch_size: 128
+      lr: 5e-4
+      weight_decay: 5e-2
+  - path: finetune_main.py
+    description: Downstream fine-tuning entry point
+  - path: sh/finetune_CSBrain_BCI2a.sh
+    description: Example fine-tuning shell script
+  models:
+  - path: models/CSBrain.py
+    description: Core CSBrain model architecture
+  - path: models/CSBrain_transformer.py
+    description: Transformer layers with CST and SSA modules
+  - path: models/model_for_bciciv2a.py
+    description: BCIC-IV-2a specific model wrapper with brain region encoding
+datasets_supported:
+- BCIC-IV-2a
+- BCIC-IV-2b
+- CHB-MIT
+- TUAB
+- TUEV
+- SEED-V
+- SEED-VIG
+- ISRUC-Sleep
+- PhysioNet
+- FACED
+- Mumtaz2016
+- SHU
+- Siena
+- HMC
+- TUSL
+- Stress
+key_innovations:
+- Cross-scale Spatiotemporal Tokenization (CST): Aggregates multi-scale features within
+    temporal windows and brain regions
+- Structured Sparse Attention (SSA): Captures long-range dependencies while eliminating
+    spurious correlations
+- Masked autoencoding pretraining for generalizable EEG representations
+- Brain region encoding (frontal, parietal, temporal, occipital, central) for spatial
+  awareness

package/AISB/catalog/aisb.t3.055_csbrain.zh.yaml

@@ -0,0 +1,184 @@
+id: aisb.t3.055_csbrain
+name: 'CSBrain: 跨尺度时空脑基础模型用于脑电解码'
+version: 0.1.0
+one_line: 跨尺度脑电基础模型，采用CST分词和SSA注意力机制，支持跨16个数据集的通用脑解码。
+task_description: |-
+  CSBrain是一个用于脑电解码的跨尺度时空脑基础模型，在NeurIPS 2025发表。该模型针对现有脑电基础模型采用尺度无关密集建模的局限性，引入了两个关键组件：(i) 跨尺度时空分词（CST），它将局部时间窗口和解剖学脑区域内的多尺度脑电特征聚合为紧凑的尺度感知分词表示；(ii) 结构化稀疏注意力（SSA），它在最小化噪声脑电信号常见虚假相关性的同时，捕获时间窗口和脑区域之间的长距离依赖关系。CST和SSA交替堆叠，以逐步整合跨尺度时空依赖关系。该基准支持从原始脑电信号进行预训练，以及跨16个数据集的下游微调，涵盖11种代表性脑电任务，包括情绪识别、运动 imagery、癫痫检测和睡眠分期。环境设置委托给CBraMod框架；预训练权重可通过Google Drive获取。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 1d+
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- balanced_accuracy
+- cohens_kappa
+- weighted_f1
+- eeg_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: |-
+    使用提供的微调权重在单个下游数据集上运行评估。加载预训练的CSBrain检查点，并在16个支持的数据集之一（如BCIC-IV-2a、CHB-MIT、TUAB、SEED-V）上进行评估。收集指标的最快路径。
+- id: pretrain_and_finetune
+  label: 预训练 + 微调
+  description: |-
+    完整流程：从头开始使用掩码自编码在大规模脑电数据上进行CSBrain预训练，然后在下游数据集上进行微调。支持16个数据集，包括BCIC-IV-2a、Siena、HMC、TUSL和所有CBraMod基准。默认训练使用40个epoch、批次大小128、学习率5e-4。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: CBraMod基准套件
+    type: benchmark_package
+    notes: 支持16个下游数据集，包括BCIC-IV-2a、BCIC-IV-2b、CHB-MIT、TUAB、TUEV、SEED-V、SEED-VIG、ISRUC-Sleep、PhysioNet、FACED、Mumtaz2016、SHU、Siena、HMC、TUSL和Stress。
+  - name: 预训练数据集
+    type: external
+    notes: 按照CBraMod预处理流程进行预训练数据准备。
+  - name: Google Drive预训练权重
+    type: external
+    url: https://drive.google.com/drive/folders/1-GsVVewRM0B93H08yts5m53yU2whxYvj
+    notes: CSBrain.pth包含预训练基础模型权重；pth_downtasks/包含所有16个下游数据集的微调权重。推荐用于无需重训练的评估场景。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 下载权重可能需要Google Drive身份验证。
+  - 需要访问CBraMod框架以完成完整数据集流程。
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 仅评估时单个GPU即可；预训练需要更多资源。
+  recommended:
+    cpu_cores: 64
+    ram_gb: 256
+    disk_gb: 500
+    gpu_count: 4
+    gpu_vram_gb: 80
+    notes: 推荐使用多GPU设置进行预训练（支持pretrain_main.py中的parallel=True）。单个数据集微调可在单块A100 80GB上运行；完整多数据集评估受益于4×80GB配置。
+environment:
+  python: 3.8+
+  cuda: 11.7+
+  pytorch: 2.0+
+  flash_attn: null
+  key_packages:
+  - torch
+  - numpy
+  - scikit-learn
+  - CBraMod
+  notes:
+  - 环境设置完全委托给CBraMod框架，遵循捆绑的README说明。
+  - 首先安装CBraMod依赖项，然后安装CSBrain特定需求（如果有）。
+  - preprocessing/文件夹中提供了Siena、HMC和TUSL数据集的预处理脚本。
+  - 无独立requirements.txt；依赖解析通过CBraMod处理。
+risk_flags:
+- large_pretraining_cost
+- multi_dataset_evaluation_overhead
+- google_drive_access
+risk_notes:
+- 从头开始预训练CSBrain计算成本高昂；对于大多数评估场景，推荐下载预训练权重。
+- 在所有16个数据集上运行完整微调需要大量磁盘空间和时间（推荐硬件上预计2-4天）。
+- Google Drive权重可能需要手动身份验证或VPN访问（取决于地理位置）。
+- CBraMod依赖可能与同一环境中的其他基准产生版本冲突。
+recommended_when: 当您需要一个具有广泛脑电迁移学习评估的脑信号基础模型基准时使用此基准。非常适合研究跨尺度表示、结构化稀疏注意力或跨异构脑电数据集的泛化能力。适用于医疗AI研究、脑机接口开发和科学机器学习探索。
+not_recommended_when: 如果您需要轻量级基准、GPU资源有限（<24GB显存）或无法托管多个脑电数据集和检查点（推荐500GB+），请勿使用此基准。不适合缺乏CBraMod框架支持或Google Drive访问受限的环境。
+paper:
+  title: 'CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding'
+  venue: NeurIPS 2025 Spotlight
+  year: 2025
+  authors:
+  - Yuchen Zhou
+  - Jiamin Wu
+  - Zichen Ren
+  - Zhouheng Yao
+  - Weiheng Lu
+  - Kunyu Peng
+  - Qihao Zheng
+  - Chunfeng Song
+  - Wanli Ouyang
+  - Chao Gou
+  url: https://arxiv.org/abs/2506.23075
+  github: https://github.com/yuchen2199/CSBrain
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.055_csbrain.zip
+  archive_type: zip
+  local_dir_name: paper-55-CSBrain
+  notes:
+  - 包含sh/文件夹，包含所有16个数据集的微调脚本。
+  - 预训练权重（CSBrain.pth）和下游权重（pth_downtasks/）需从Google Drive单独下载。
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.055_csbrain.zip
+  sha256: 3d02943b88e62e967308b3cc9bae379cab2431139fcfef0f591fbc95fb206d4c
+  size_bytes: 123771
+display:
+  palette_seed: violet-cyan-neuro
+  art_style: brain-atlas
+  accent_priority: high
+image_path: ../image/055_aisb.t3.055_csbrain.jpg
+track_fit:
+- paper_track
+- benchmark_track
+capability_tags:
+- research_code_optimization
+- eeg_decoding
+- brain_foundation_models
+- scientific_ml
+- healthcare_ai
+aisb_direction: T3
+code_anchors:
+  evaluation:
+  - path: evaluate_bci2a_v2.py
+    description: BCIC-IV-2a数据集的主要评估脚本
+    metrics:
+    - balanced_accuracy
+    - cohens_kappa
+    - weighted_f1
+  training:
+  - path: pretrain_main.py
+    description: 基础模型预训练入口点
+    defaults:
+      epochs: 40
+      batch_size: 128
+      lr: 5e-4
+      weight_decay: 5e-2
+  - path: finetune_main.py
+    description: 下游微调入口点
+  - path: sh/finetune_CSBrain_BCI2a.sh
+    description: 微调示例shell脚本
+  models:
+  - path: models/CSBrain.py
+    description: 核心CSBrain模型架构
+  - path: models/CSBrain_transformer.py
+    description: 带有CST和SSA模块的Transformer层
+  - path: models/model_for_bciciv2a.py
+    description: 带有脑区域编码的BCIC-IV-2a专用模型包装器
+datasets_supported:
+- BCIC-IV-2a
+- BCIC-IV-2b
+- CHB-MIT
+- TUAB
+- TUEV
+- SEED-V
+- SEED-VIG
+- ISRUC-Sleep
+- PhysioNet
+- FACED
+- Mumtaz2016
+- SHU
+- Siena
+- HMC
+- TUSL
+- Stress
+key_innovations:
+- 跨尺度时空分词（CST）：在时间窗口和脑区域内聚合多尺度特征
+- 结构化稀疏注意力（SSA）：捕获长距离依赖关系，同时消除虚假相关性
+- 用于通用脑电表示的掩码自编码预训练
+- 脑区域编码（额叶、顶叶、颞叶、枕叶、中央）以实现空间感知