知识卡片
测试偏差会扭曲模型选择
测试集决定选择哪种算法和训练数据;若它偏向某种场景,更有代表性的训练集也可能在该测试上输给同样有偏的来源。评估目标应与部署条件对应,并保留跨来源、跨场景的检查,避免把对单一基准的适应当作真实改进。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
One usually thinks about the biases in the training set, but biases in the test set can have an even more negative impact on the final system behavior than the training bias. The test set is what researchers and engineers use to decide which algorithm is best, or what data collection is more effective for training. However, there might be little correlation between the performance
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第35章《数据偏差与分布漂移》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c035b.xhtml)
- 结论依据:摘录中关于测试偏差会扭曲模型选择的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:One usually thinks about the biases in the training set, but biases in the test set can have an even more negative impact on the final system behavior than the training bias. The test set is what researchers and engineers use to decide which algorithm is best, or what data collection is more effective for training. However, there might be little correlation between the performance