Home首页 · Machine Learning机器学习

Machine Learning机器学习

Four labs. The first two use this year’s trust score in 24 districts and the share of households in public rental housing: first how much the curve may bend, then a pull that shortens or deletes slopes. The third is 48 students, writing-workshop hours, and grades, with one line or four. The fourth is 24 consultation forms from last year and 12 street interviews from this week.四个实验。前两个用今年 24 个区的信任分数和公屋住户占比：先看曲线可以弯多少，再看把斜率缩短或删掉的拉力。第三个是 48 名学生、写作工作坊时数和成绩，一条线或四条。第四个是去年 24 份咨询表和本周 12 份街访。

A curve can look right on the 24 districts used to draw it and miss a new set of districts drawn the same way. The two numbers under the chart are that miss this year and the miss next year.一条曲线可以在用来画它的 24 个区上看起来对，却错过按同样方式再抽的一批区。图下的两个数，是今年的偏差和明年的偏差。

A pitfall.一个陷阱。 Next year’s miss on the screen was watched while the bend or the pull was chosen. It is not a year the office has not yet looked at.屏幕上明年的偏差，是在挑选弯曲或拉力时已经看过的。它不是办公室还没看过的那一年。

Labs实验

• Lab实验 Overfitting过拟合 Will this year’s curve still work on next year’s districts?今年的曲线在明年的区上还管用吗？
• Lab实验 Regularization正则化 What happens when slopes are shortened or deleted?斜率被缩短或删掉时，会发生什么？
• Glance一瞥 Group slopes分组斜率 One line for the university, or one line per faculty?全校一条线，还是每个学院一条？
• Glance一瞥 Updating a belief更新一个看法 Last year’s notes, a street survey, and where the briefing moves.去年的笔记、一次街访，以及简报移到哪里。

Related paths相关路径

Overfitting journey过拟合之旅 walks a straight line, a gentle curve, and a curve that bends to every district.依次走一条直线、一条缓弯，和一条弯向每一个区的曲线。Shrink the fit把拟合收一收 continues from that bend into a pull that shortens slopes, then one that can delete them.从那个弯曲继续走到缩短斜率的拉力，再走到可以删掉斜率的拉力。