Distribution shift analysis
Model performance is compared between training data and deployment data to detect shifts in inputs or labels that reduce accuracy, with monitoring set up to catch drift.
GCRI · 人工智能与自主系统
Artificial intelligence and machine learning is the technology this platform stewards for the network. Here it is studied as a technology whose behaviour must be understood: how models learn from data, generalise, fail outside their training distribution and are adapted to specific uses.
Artificial intelligence and machine learning develop algorithms that learn patterns from data to classify, predict or generate outputs. They are increasingly used for hazard detection, forecasting and decision support, and are themselves a source of risk. In artificial intelligence and data and compute, machine learning can scale detection and forecasting across interconnected 系统. Its own failure modes must be weighed when automated outputs inform consequential decisions. The canonical description is on STEM Commons.
Model performance is compared between training data and deployment data to detect shifts in inputs or labels that reduce accuracy, with monitoring set up to catch drift.
General models are adapted to a narrow task with curated data, and the adapted model is evaluated for both task performance and retained safety behaviour.
Feature attribution and probing methods are used to examine which inputs drive a model’s outputs, with their limits as explanations stated.
每个主体在足够多的联系人行动后才行动;平均阈值或早期行动者的微小变化,决定该行为是停留在局部还是在整个人群中级联扩散。这是带有说明性取值的教学模型,不是预测。
控制措施
读数
| 步骤 | 已采取行动的比例(%) |
|---|---|
| 0 | 2.3 |
| 1 | 6.5 |
| 2 | 16.0 |
| 3 | 28.5 |
| 4 | 47.0 |
| 5 | 65.3 |
| 6 | 80.3 |
| 7 | 93.0 |
| 8 | 99.3 |
| 9 | 100.0 |
| 10 | 100.0 |
No observed data enters this model: it runs on the teaching values set by the controls, to be replaced with a programme’s own evidence.
贯穿六大支柱,由 Nexus Consortium 牵头
| 数据 | 角色 | 标签 |
|---|---|---|
| Training and validation datasets with documentation | Model development | 已声明 |
| Deployment input logs | Shift detection | 观测 |
| Benchmark suites | 比较 | 已报告 |
模型、其假设、数据和不确定性,以科学家或工程师会核查的形式呈现。结构遵循已发表的方法。所取数值仅为示意,不是研究发现。
The model structure used for machine learning behaviour, the assumptions it makes and how each is tested.
Ri = f(Hi, Ei, Vi, Ci)危害、风险暴露、脆弱性和应对能力EAL = ∑k pk · Lk对年概率为 pkxt+1 = A xt + B ut + εtA carries the couplings, u the shock, ε the error| 假设 | 为何制作 | 如何检验 |
|---|---|---|
| 在一次事件中,危害与暴露相互独立 | 使损失积分保持可计算 | 与基于联合记录拟合的 copula 进行比较 |
| 耦合在工作点附近是线性的 | 让冲击通过 A 传播 | 在线性范围之外施加扰动并比较 |
| 脆弱性曲线可以在不同地点之间迁移使用 | 很少有本地曲线可用 | 在有观测损失的地方据此进行验证 |
模型在尊重空间、时间与群体结构的留出数据上进行评估。在测试准确性的同时,也测试稳健性、校准与公平性。
| 核查 | 如何开展 |
|---|---|
| Held-out evaluation | Models are evaluated on data not used for training or tuning |
| Dataset documentation | Training data carry datasheets on origin, consent and limits |
| Drift monitoring | Deployed models are monitored for input and performance drift |
当部署条件与训练数据不同时,性能会下降。模型可能固化偏见,并在罕见事件上毫无明显预警地失效。
报告 is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.
切勿发送个人数据、敏感基础设施细节或您无权分享的任何内容。贡献在使用前会经过审查;发送贡献不赋予署名权或编辑权,赞助也绝不能换取研究结论。
本网站上的地图、数据和图表仅作为背景和一般信息展示。它们不是建议、预测或预警;信息的局限见免责声明和数据说明。
This site uses only strictly necessary cookies to work and to remember your privacy choices. Nothing loads from other companies, and we never sell or share your personal information.
This site uses only strictly necessary cookies to work and to remember your privacy choices. Nothing loads from other companies, and we never sell or share your personal information. We run no advertising. You can review your choices at any time through Privacy choices at the foot of every page.