@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.100_invmissingdata.zh.yaml

@@ -0,0 +1,150 @@
+id: aisb.t3.100_invmissingdata
+name: 核心点插补 (KPI)
+version: 0.1.0
+one_line: 双层内核自适应迭代插补利用预言机功能监督进行异构丢失数据恢复。
+task_description: 该基准测试实现了 KPI（核点插补），这是一种用于迭代缺失数据插补的双层优化框架。内循环通过再生内核希尔伯特空间（RKHS）中的功能优化来适应每个特征的插补模型，解决异构数据生成过程中的模型错误指定问题。外循环利用预言机（完全观察）功能作为特权监督信号。支持仅 CPU 执行最小路由； GPU 加速可用于整个工作流程。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- mae
+- wass
+- rmse
+- imputation_report
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 使用默认超参数在准备好的数据集上运行打包的 KPI 基准脚本。适用于从预处理的 UCI 数据中快速提取指标，无需经过全面培训。
+- id: full_train_eval
+  label: 完整训练+评估
+  description: 执行完整的 KPI 管道：内核点插补优化、具有 Oracle 功能监督的双层训练以及跨所有支持的数据集的完整基准测试套件。包括 Wasserstein 距离评估。
+dataset_download:
+  primary_method: manual
+  sources:
+  - name: UCI Machine Learning Repository
+    url: https://archive.ics.uci.edu
+    notes: 'Pre-processed tabular datasets (e.g., blood_transfusion, concrete_compression) should be placed in the ./dataset folder. Specific dataset references are documented in benchmark_kpi.py argument defaults and aisb_catalog.yaml.
+
+      '
+  - name: Benchmark archive
+    url: https://deepscientist.cc/AISB/100_invmissingdata
+    archive_type: zip
+    local_dir_name: paper-100-InvMissingData
+  notes:
+  - 需要 UCI 源数据和打包的基准档案。
+  - 捆绑的自述文件中详细介绍了数据集准备步骤。
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 30
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: 最低配置完全支持仅 CPU 执行。
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 8
+    notes: GPU推荐使用full_train_eval来加速内核计算和双层优化。
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.7.1
+  flash_attn: null
+  key_packages:
+  - ot (Optimal Transport)
+  - numpy
+  - pandas
+  - scipy
+  - torch
+  notes:
+  - 对于最短路径来说，仅 CPU 执行是合理的。
+  - 请参阅捆绑的 requirements.txt 以获取完整的依赖项集。
+  - 提供了可选的 Docker 容器，其具有名为“kpi”的基于 conda 的环境。
+risk_flags: []
+risk_notes:
+- 静态代码审核确认所有阶段指标都存在可执行锚点。
+- 在此打包过程中未执行任何运行时执行；指标值需要通过执行进行验证。
+recommended_when: 当您需要一个缺失数据基准测试时，请使用此基准测试，该基准测试专注于灵活的非参数插补行为以及内核适应几何和预言功能监督。适用于具有异构缺失模式的表格数据。
+not_recommended_when: 如果您需要大量的多模式基准测试或基于深度学习的插补而没有明确的内核方法基础，请不要使用此选项。
+paper:
+  title: Inverse Methods for Missing Data Imputation
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/kim25ag.html
+  code_url: null
+  baseline_id: aisb.t3.100_invmissingdata
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.100_invmissingdata.zip
+  archive_type: zip
+  local_dir_name: paper-100-InvMissingData
+  notes:
+  - Archive contains benchmark_kpi.py, model/KPI.py, scripts/, and requirements.txt.
+  - UCI datasets must be obtained separately from archive.ics.uci.edu.
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.100_invmissingdata.zip
+  sha256: 0b54fbad4ac73321054a33a6dc80e069c22a81cf54788c237d938a8f1caf70ef
+  size_bytes: 1388561
+display:
+  palette_seed: bone-blue-impute
+  art_style: imputation-lab
+  accent_priority: medium
+  tags:
+  - research_code_optimization
+  - missing_data
+  - imputation
+  - kernel_methods
+  - tabular_ml
+image_path: ../image/100_aisb.t3.100_invmissingdata.jpg
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3
+capability_tags:
+- research_code_optimization
+- missing_data
+- imputation
+- kernel_methods
+- tabular_ml
+executable_anchors:
+  benchmark_script: benchmark_kpi.py
+  core_model: model/KPI.py
+  scripts_dir: scripts/
+  evaluation_protocol:
+  - scripts/qb.sh (quick_eval)
+  - scripts/is.sh (full_train_eval)
+metric_evidence:
+  mae:
+    origin_path: model/remasker/model_mae.py
+    source_ref: code-backed
+    status: executable_anchor_found
+  wass:
+    origin_path: benchmark_kpi.py
+    source_ref: code-backed
+    status: executable_anchor_found
+  rmse:
+    origin_path: benchmark_kpi.py
+    source_ref: code-backed
+    status: executable_anchor_found
+  mse:
+    origin_path: benchmark_kpi.py
+    source_ref: code-backed
+    status: executable_anchor_found
+caveat_snapshot: 'Static code audit confirmed executable anchors for every staged metric. Runtime execution has not been performed; metric values require verification before trust.
+
+  '

package/AISB/catalog/aisb.t3.101_acia.yaml

@@ -0,0 +1,164 @@
+id: aisb.t3.101_acia
+name: Measure-Theoretic Anti-Causal Representation Learning
+version: 0.1.0
+one_line: Two-level causal representation learning via measure-theoretic intervention
+  kernels for anti-causal OOD generalization across multi-environment settings.
+task_description: 'This packaged benchmark implements ACIA (Anti-Causal Invariant
+  Abstractions), a measure-theoretic framework for learning causally invariant representations
+  in the anti-causal setting where labels Y cause observations X. The method employs
+  a two-level architecture: low-level representations (phi_L) capture how labels generate
+  observable features including spurious environment-specific correlations, while
+  high-level representations (phi_H) distill environment-invariant causal patterns.
+  ACIA handles both perfect and imperfect interventions via interventional kernels
+  without requiring explicit structural causal models. Training optimizes a composite
+  loss combining standard cross-entropy with R1 (environment independence) and R2
+  (intervention invariance) regularizers, with optional R3 (feature suppression) for
+  color-removal scenarios. Evaluation includes accuracy on held-out environments,
+  environment_invariance metrics measuring representation stability across domains,
+  low_level_invariance tracking phi_L robustness, and intervention_robustness measuring
+  performance under do(X) and do(Y) interventions.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- accuracy
+- environment_invariance
+- low_level_invariance
+- intervention_robustness
+launch_profiles:
+- id: quick_cmnist
+  label: Quick CMNIST
+  description: 'Run the optimized fast CMNIST route using vectorized dataset construction.
+    Downloads MNIST via torchvision and automatically recolors into e1/e2 variants.
+    Reports accuracy and invariance metrics in a single training run (~10 epochs).
+    Suitable for rapid validation and hyperparameter tuning.
+
+    '
+- id: paper_cmnist
+  label: Paper CMNIST
+  description: 'Execute the full paper-style perfect-intervention training route matching
+    Table 1 evaluation conditions. Includes both environment training loops, comprehensive
+    metric logging, and intervention robustness testing. Matches the 97.7% CMNIST
+    accuracy reported in the paper.
+
+    '
+dataset_download:
+  primary_method: torchvision
+  sources:
+  - name: MNIST
+    url: torchvision.datasets.MNIST
+    auto_download: true
+    notes: 'MNIST downloads automatically via torchvision when dataset is first instantiated.
+      Recoloring into ColoredMNIST environment variants (e1: even→red, e2: even→green)
+      happens locally in-memory. No manual staging required.
+
+      '
+  notes:
+  - CMNIST synthetic benchmark requires no external data staging beyond torchvision
+    MNIST.
+  - Real-world datasets (Camelyon17) require additional download steps documented
+    in examples/04_camelyon17.py.
+  - Paper reports 84.40% on Camelyon17 with 19% improvement over LECI baseline.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No external API credentials, dataset accounts, or HPC allocations required for
+    CMNIST path.
+  - The packaged CMNIST route uses only locally cached MNIST data.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 30
+    gpu_count: 1
+    gpu_vram_gb: 8
+    notes: 'GPU required for reasonable training throughput; CPU-only training feasible
+      but significantly slower (~5-10x). Single GPU suffices due to CMNIST scale.
+
+      '
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: '16GB VRAM accommodates full batch sizes (128) for CMNIST without memory
+      pressure. Additional RAM supports larger custom datasets if extending beyond
+      packaged benchmarks.
+
+      '
+environment:
+  python: '3.8'
+  cuda: '11.3'
+  pytorch: 1.12.0
+  flash_attn: null
+  key_packages:
+  - torch>=1.12.0
+  - torchvision>=0.13.0
+  - numpy
+  notes:
+  - Python 3.8+ required; tested with 1.12.0 but compatible with newer PyTorch versions.
+  - CUDA 11.3+ recommended for GPU training; CPU fallback supported for debugging.
+  - Full dependency set in requirements.txt; run pip install -e . for editable install
+    with all extras.
+risk_flags: []
+risk_notes:
+- No runtime file I/O to external storage beyond local MNIST cache.
+- No network calls during training; MNIST cached after first download.
+- Training is deterministic with fixed seeds (42) in packaged scripts.
+recommended_when: 'Use this benchmark when you need to evaluate causal invariance
+  in multi-environment settings with explicit environment labels. Ideal for benchmarking
+  anti-causal learning methods, testing intervention robustness, or comparing representation
+  learning approaches on structured synthetic data with known ground-truth causal
+  structure.
+
+  '
+not_recommended_when: 'Do not use this benchmark for large-scale multimodal tasks,
+  generative modeling, or benchmarks requiring multi-GPU training. Not suitable when
+  explicit SCM/DAG knowledge is available and preferred—ACIA specifically avoids SCM
+  requirements.
+
+  '
+paper:
+  title: Measure-Theoretic Anti-Causal Representation Learning
+  venue: NeurIPS
+  year: 2025
+  url: https://openreview.net/forum?id=lp6XOoQ4VL
+  authors:
+  - Arman Behnam
+  - Binghui Wang
+  institution: Illinois Institute of Technology
+  notes: 'Paper introduces ACIA framework with theoretical bounds on OOD generalization.
+    Claims 97.7% CMNIST accuracy and 84.40% on Camelyon17 with 19% improvement over
+    LECI. Static audit confirms executable anchors for all reported metrics.
+
+    '
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.101_acia.zip
+  archive_type: zip
+  local_dir_name: paper-101-ACIA
+  notes: 'Download contains full source with acia/ package, examples/, run_*.py scripts,
+    and json/metric_contract.json for metric specification.
+
+    '
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.101_acia.zip
+  sha256: 12590ff23562c80d2cf91dc224d44ba1909950585248f0717624a9914a2fb89a
+  size_bytes: 3127466
+display:
+  palette_seed: violet-sand-invariant
+  art_style: causal-abstraction
+  accent_priority: medium
+image_path: ../image/101_aisb.t3.101_acia.jpg

package/AISB/catalog/aisb.t3.101_acia.zh.yaml

@@ -0,0 +1,109 @@
+id: aisb.t3.101_acia
+name: 测度论反因果表征学习
+version: 0.1.0
+one_line: 通过测量理论干预内核进行两级因果表示学习，以实现跨多环境设置的反因果 OOD 泛化。
+task_description: 该打包基准实现了 ACIA（反因果不变抽象），这是一个测量理论框架，用于在反因果设置中学习因果不变表示，其中标签 Y 引起观察 X。该方法采用两级架构：低级表示 (phi_L) 捕获标签如何生成可观察特征，包括虚假的环境特定相关性，而高级表示 (phi_H) 提取环境不变因果模式。 ACIA 通过干预内核处理完美和不完美的干预，而不需要明确的结构因果模型。训练将标准交叉熵与 R1（环境独立性）和 R2（干预不变性）正则化器相结合，以及针对颜色去除场景的可选 R3（特征抑制）来优化复合损失。评估包括保留环境的准确性、测量跨域表示稳定性的 environment_invariance 指标、low_level_invariance 跟踪 phi_L 稳健性以及 intervention_robustness 测量 do(X) 和 do(Y) 干预下的性能。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- accuracy
+- environment_invariance
+- low_level_invariance
+- intervention_robustness
+launch_profiles:
+- id: quick_cmnist
+  label: 快速CMNIST
+  description: 使用矢量化数据集构建运行优化的快速 CMNIST 路线。通过 torchvision 下载 MNIST 并自动重新着色为 e1/e2 变体。报告单次训练运行（约 10 个时期）的准确性和不变性指标。适用于快速验证和超参数调整。
+- id: paper_cmnist
+  label: 论文CMNIST
+  description: 执行全纸式完美干预训练路线匹配表1评价条件。包括环境训练循环、综合指标记录和干预鲁棒性测试。与论文中报告的 97.7% CMNIST 准确率相符。
+dataset_download:
+  primary_method: torchvision
+  sources:
+  - name: MNIST
+    url: torchvision.datasets.MNIST
+    auto_download: true
+    notes: 'MNIST downloads automatically via torchvision when dataset is first instantiated. Recoloring into ColoredMNIST environment variants (e1: even→red, e2: even→green) happens locally in-memory. No manual staging required.
+
+      '
+  notes:
+  - CMNIST 综合基准测试不需要除 torchvision MNIST 之外的外部数据暂存。
+  - 真实世界数据集 (Camelyon17) 需要示例/04_camelyon17.py 中记录的额外下载步骤。
+  - 论文报告 Camelyon17 的准确率为 84.40%，比 LECI 基线提高了 19%。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - CMNIST 路径不需要外部 API 凭据、数据集帐户或 HPC 分配。
+  - 打包的CMNIST路由仅使用本地缓存的MNIST数据。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 30
+    gpu_count: 1
+    gpu_vram_gb: 8
+    notes: 合理的训练吞吐量所需的 GPU；仅 CPU 训练可行，但速度明显慢（约 5-10 倍）。由于 CMNIST 规模，单个 GPU 就足够了。
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: 16GB VRAM 可容纳 CMNIST 的完整批量大小 (128)，且无内存压力。如果超出打包基准测试范围，额外的 RAM 可支持更大的自定义数据集。
+environment:
+  python: '3.8'
+  cuda: '11.3'
+  pytorch: 1.12.0
+  flash_attn: null
+  key_packages:
+  - torch>=1.12.0
+  - torchvision>=0.13.0
+  - numpy
+  notes:
+  - 需要 Python 3.8+；使用 1.12.0 进行测试，但与较新的 PyTorch 版本兼容。
+  - GPU 训练推荐使用 CUDA 11.3+；支持 CPU 回退以进行调试。
+  - requirements.txt 中设置的完整依赖项；运行 pip install -e 。用于包含所有附加功能的可编辑安装。
+risk_flags: []
+risk_notes:
+- 除了本地 MNIST 缓存之外，没有运行时文件 I/O 到外部存储。
+- 训练期间无网络调用；首次下载后缓存 MNIST。
+- 训练是确定性的，使用打包脚本中的固定种子 (42)。
+recommended_when: 当您需要评估具有显式环境标签的多环境设置中的因果不变性时，请使用此基准。非常适合对反因果学习方法进行基准测试、测试干预稳健性或将结构化合成数据的表示学习方法与已知的真实因果结构进行比较。
+not_recommended_when: 请勿将此基准测试用于大规模多模式任务、生成建模或需要多 GPU 训练的基准测试。当明确的 SCM/DAG 知识可用且首选时不适合 — ACIA 特别避免了 SCM 要求。
+paper:
+  title: Measure-Theoretic Anti-Causal Representation Learning
+  venue: NeurIPS
+  year: 2025
+  url: https://openreview.net/forum?id=lp6XOoQ4VL
+  authors:
+  - Arman Behnam
+  - Binghui Wang
+  institution: Illinois Institute of Technology
+  notes: 论文介绍了 ACIA 框架以及 OOD 泛化的理论界限。声称 CMNIST 准确率达到 97.7%，Camelyon17 准确率达到 84.40%，比 LECI 提高 19%。静态审计确认所有报告指标的可执行锚点。
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.101_acia.zip
+  archive_type: zip
+  local_dir_name: paper-101-ACIA
+  notes: 'Download contains full source with acia/ package, examples/, run_*.py scripts, and json/metric_contract.json for metric specification.
+
+    '
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.101_acia.zip
+  sha256: 12590ff23562c80d2cf91dc224d44ba1909950585248f0717624a9914a2fb89a
+  size_bytes: 3127466
+display:
+  palette_seed: violet-sand-invariant
+  art_style: causal-abstraction
+  accent_priority: medium
+image_path: ../image/101_aisb.t3.101_acia.jpg

package/AISB/catalog/aisb.t3.102_stochasticff.yaml

@@ -0,0 +1,178 @@
+id: aisb.t3.102_stochasticff
+name: Stochastic Forward-Forward Learning through Representational Dimensionality
+  Compression
+version: 0.1.0
+one_line: Non-BP training via effective dimensionality goodness function that compresses
+  within-class representations and leverages stochasticity as a computational resource.
+task_description: 'This benchmark implements stochastic forward-forward (FF) learning
+  with a novel "dimensionality compression" goodness function derived from effective
+  dimensionality (ED) of neural responses. Unlike standard FF that requires hand-crafted
+  negative samples and uses sum-of-squared-activations as goodness, this approach
+  exploits second-order statistical structure through the eigenvalue spectrum of uncentered
+  covariance matrices. Training objectives minimize ED for noisy isologues of individual
+  inputs while maximizing ED across the sample distribution, eliminating the need
+  for explicit negative sample generation. The method achieves competitive accuracy
+  against other non-backpropagation methods on CIFAR-10, CIFAR-100, and MNIST, with
+  inherent support for energy-based inference from mean-squared outputs. Includes
+  information breakdown analysis tools for estimating mutual information in individual
+  neurons and synergistic neuron combinations.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- cifar10_val_accuracy
+- cifar100_val_accuracy
+- mnist_val_accuracy
+- best_model_checkpoint
+- training_log
+- representation_analysis
+launch_profiles:
+- id: quick_cifar10
+  label: Quick CIFAR-10
+  description: 'Execute the CIFAR-10 training route with default hyperparameters.
+    Downloads datasets via torchvision, caches ZCA statistics under ./data/, saves
+    checkpoints to ./checkpoint/, and reports final validation accuracy. Typical runtime
+    2-4 hours on single GPU.
+
+    '
+  hardware_notes: 'Default config uses 3 epochs, batch size 128, learning rate 0.001
+    with cosine annealing.
+
+    '
+- id: full_suite_analysis
+  label: Full Suite + Analysis
+  description: 'Run multi-dataset training workflow across CIFAR-10, CIFAR-100, and
+    MNIST, then invoke bundled analysis route to produce information breakdown plots
+    (Fig. 4c/5c equivalents). Requires manually calling specific experiment functions
+    per training.py comments.
+
+    '
+  hardware_notes: 'Full suite with analysis extends runtime proportionally. Phase
+    2 training uses 60 epochs.
+
+    '
+dataset_download:
+  primary_method: torchvision
+  sources:
+  - name: CIFAR-10
+    path: torchvision.datasets.CIFAR10
+  - name: CIFAR-100
+    path: torchvision.datasets.CIFAR100
+  - name: MNIST
+    path: torchvision.datasets.MNIST
+  notes:
+  - Datasets are downloaded automatically on first run and cached locally under ./data/
+  - ZCA whitening statistics are precomputed and cached per dataset under ./data/
+  - The cited fast_mnist snippet informs helper implementation; it is not the dataset
+    acquisition route
+  - No external API credentials required for torchvision-served datasets
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No explicit dataset or API credentials detected in bundled source
+  - GitHub repository available at https://github.com/ZhichaoZhu/StochasticForwardForward
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 30
+    gpu_count: 1
+    gpu_vram_gb: 8
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+environment:
+  python: '3.9'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - torch>=2.0
+  - torchvision
+  - numpy
+  - scipy
+  - scikit-learn
+  - pandas
+  - tqdm
+  notes:
+  - See bundled requirements.txt for pinned versions
+  - See bundled README/requirements for full dependency set
+risk_flags:
+- experimental_algorithm
+- nonstandard_training_paradigm
+risk_notes:
+- FF learning is less mature than backpropagation; results may vary across random
+  seeds
+- No standard fine-tuning recipes exist for this algorithm family
+- Hyperparameter sensitivity: ED loss balance factors, sampling length, projecting
+    dimensions require tuning
+recommended_when: '- Exploring alternative training paradigms to backpropagation -
+  Studying biologically plausible learning algorithms - Investigating neuromorphic
+  computing applications where stochasticity is a resource - Research on representation
+  learning with second-order statistical structure - Information-theoretic analysis
+  of neural network representations
+
+  '
+not_recommended_when: '- Requiring conventional large-model finetuning workflows -
+  Needing production-grade training pipelines with extensive tooling - Tasks requiring
+  rapid iteration without algorithm-specific expertise
+
+  '
+paper:
+  title: Stochastic Forward-Forward Learning through Representational Dimensionality
+    Compression
+  authors:
+  - Zhu, Zhichao
+  - Qi, Yang
+  - Ma, Hengyuan
+  - Lu, Wenlian
+  - Feng, Jianfeng
+  venue: NeurIPS
+  year: 2025
+  url: https://arxiv.org/abs/2505.16649
+  github: https://github.com/ZhichaoZhu/StochasticForwardForward
+  citation: "@article{zhu2025stochastic,\n  title={Stochastic Forward-Forward Learning\
+    \ through Representational Dimensionality Compression},\n  author={Zhu, Zhichao\
+    \ and Qi, Yang and Ma, Hengyuan and Lu, Wenlian and Feng, Jianfeng},\n  journal={arXiv\
+    \ preprint arXiv:2505.16649},\n  year={2025}\n}\n"
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.102_stochasticff.zip
+  archive_type: zip
+  local_dir_name: paper-102-StochasticFF
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.102_stochasticff.zip
+  sha256: 4eab3f39d4cc098ee47e9d4a31c8485f3b09752123b134a05292f5bfababe41e
+  size_bytes: 65412
+display:
+  palette_seed: amber-indigo-forward
+  art_style: learning-dynamics
+  accent_priority: medium
+image_path: ../image/102_aisb.t3.102_stochasticff.jpg
+capability_tags:
+- forward_forward_learning
+- representation_learning
+- non_backpropagation
+- effective_dimensionality
+- image_classification
+- efficient_training
+- biologically_plausible_learning
+- information_breakdown_analysis
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3