引導式測驗 · 中等 等級
醫療保健測驗中的聯邦學習
測試分散式訓練、原始資料局部性、隱私限制、站點異質性、外部評估和治理要求。
問題 1 的 8
What remains at participating sites in federated learning?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What remains at participating sites in federated learning?
- No data or updates are transmitted
- Only predictions are shared with patients
- Raw data can remain local while model updates are shared
- Every patient record moves to one central database
- Does federated learning automatically guarantee privacy?
- No; updates and metadata may still leak information
- Yes, because raw data never move
- Yes, any model update is anonymous
- No privacy controls can be used
- Why align feature and outcome definitions across hospitals?
- To make every hospital use the same vendor
- 消除驗證的需要
- To ensure each site has identical patients
- Different coding and measurement can create inconsistent training signals
- Which evaluation best tests generalization for a federated model?
- Measuring communication speed alone
- Testing at a site that did not contribute to training
- Testing only on the largest training site
- Reusing local training predictions
- How can large sites affect the federation?
- They guarantee performance for small sites
- They make privacy irrelevant
- They may dominate updates if aggregation does not account for site differences
- They automatically remove bias
- What can secure aggregation or differential privacy contribute?
- A guarantee of identical site performance
- Additional protections with possible utility trade-offs
- A guarantee that the model is clinically valid
- A replacement for access control
- What should be agreed before training begins?
- Only the color of the dashboard
- The final accuracy value
- The model name alone
- Task, data definitions, update protocol, security, and oversight responsibilities
- What does the EXAM study illustrate?
- A multi-site federated training workflow for a specific COVID-19 task
- A privacy guarantee for all medical AI
- A universal model for every disease
- A replacement for clinical evaluation