DAY 5
0

# Course - How Google does Machine Learning

• What it means to be AI first
• What it means to be AI first
• Two stages of ML
• Demo: ML in Google products
• Demo: ML in Google Photos
• Replacing heuristics
• Framing an ML problem
• Lab Intro - Framing an ML Problem
• Framing an ML problem
• Lab debrief
• ML in applications
• Demo: ML in applications
• Pre-trained models
• The ML marketplace is evolving <--- (昨天到這邊)
• `A data strategy`
• `Training and serving skew`
• `An ML strategy`
• `Transform your business`
• `Transform your business`
• `Lab Intro - ML use case`
• `Non-traditional ML use case`
• Module 2 Quiz

## 11. A data strategy (接續昨天的編號)

• What it means to be AI first
• A data strategy

wi-fi points, barometric pressure, typical walking speed 之類的資料

``````因此，ML就是能超越手寫規則的方法，

``````

(前面章節也有提到，這就是local data的問題)

If machine learning is a rocket engine, data is the fuel.

ML strategy is first and foremost a data strategy.

## 12. Training and serving skew

• What it means to be AI first
• Training and serving skew

1. 分析數據也代表著要收集資料，掌握數據是最耗時又最難的部分
2. 收集好數據還需要去rating data(finding labels)，有先分析過再進入ML階段可以避免做白工
3. 要建立一個好的ML模型，就要夠清楚你的資料，先分析數據也是為了這點
4. ML會將產品走向自動與規畫化，也就是ML是數據分析的延伸與擴展。

The problem is that the result of stream processing and the result of batch processing have to be the same.

## 13. An ML strategy

• What it means to be AI first
• An ML strategy

ML要注意的事情："ML的重點在量而不是在複雜度"

The idea is that if you're failing fast, you get the ability to iterate. This ability to experiment is critical in the realm of machine learning.

(慢慢嘗試後得到的慢速失敗(採取比較穩紮穩打的作法)，反而成功的機會會小一些。)

• What it means to be AI first
• Lab Intro - ML use case

• Did the card get canceled?
• Did the flight get delayed, etc.?
• Or did they just walk into your store and see an empty shelf?
• Why are they contacting you?
• Did your system automatically find the right action to take for this customer?
• Did it offer to rebook them?

• infuse your applications with machine learning

• use machine learning to delight your users

``````試著去思考公司內一個已經存在的應用，去想哪個部分能夠用ML取代呢?
``````
• What are some of the benefits to doing so?

• What kinds of data would you collect if you wanted to do this?

• Are you collecting that data today? If no, why not?