AIForEveryone_what is AI

AI is artificial intelligence.

There are 2 types of AI:

• ANI(Artificial Narrow Intelligence). ex) smart speaker, self-car driving, factory, etc
• AGI(Artificial General Intelligence). ex) anything human can do

Supervised learning is basically process of input and output. If emails data is input and output is selected spam mails, then this is spam-filtering AI. If input is English and output is Chinese, then this is AI translator.

Dataset is a table of data. There are several ways to collect data such as manual labeling, observing behaviors and getting data from web or partners.

In case of manual labeling, you can label the image whether the figure of the photo is cat or not. If it is cat, label it as 'Cat'. If not, 'not Cat'.

manual labeling

img1 -> cat
img2 -> not cat
img3 -> not cat
img4 -> cat

The other way of collecting data is by observation. It is possible to get housing sales or factory machine data.

time(A)	templature(A)	pressure(A)	machine fail(B)
09:00	39c	20	False
09:20	55c	40	False
09:50	60c	50	True
10:20	52	40	False
In this case, input data(A) are time, templature and pressure. Output is machine fail data. By observing what happend, you can get datas.

The other way of getting datas is from partners or websites such as data hub, api, and so on.

The quality of data matters. Garbage in Gargage out. The problems of data are normally related to two things, outliers or unknown data. For example, shoes prices are generally between $100~300. But someone could seel the shoes in $900 which is too much and not good data. The other example is when you don't know what color the shoes was. At this case, investigator put color datas as 'unknown' and this is not good data, too. To avoid bad datas pollute the data pools, AI team should do so-called cleansing work. There are many ways to cleanse.

AI is good at dealing with both structured and un-structured data. Structured data is dataset like spread sheet with well organized format. Unstructured data include image, audio, texts, etc. AI does like both of them.

Terminology

Machine learning

Imagin there is dataset about housing sales. Size of house, # of bedrooms, newlly renovated or not, price, etc. Machine Learning guys will set those datas except price as input data and price data as output to build AI system such as website or mobile app. Machine learning is used to software for predecting system.

Machine learning is the field of study that gives computers the ability to learn without being explicitly programmed. - Arthur Samuel(1959), pioneer of ML.

Data Science

In contrast, data science guys will analyze those data and create insights out of them. For example, the factor of renovation makes price n% higher. Those insights are reported to decision makers.

Data science is a science of extracting knowledges and insights from data

Deep learning

Actually, Artificial Neural Network is equal to deep learning. Deep learning is quite branded name. It is inspired by brain but the way biological brain work is a bit different.
You input datas into the box so called Neural Network which consists of so many neurons. The box is just big mathematical equation. And it returns output, in this case predicted price.

What makes an AI Company?

Internet Era

Shopping mall with website and Internet company are different. The differences are that internet company do:

• A/B testing
• iteration time is short
• dicision is made from bottom layers like engineers or domain professionals not top town.

AI Era

company with deep learning is not AI company. AI company can do :

• strategic data acquisition. - "Oh we should collect this data for x so that we can do y"
• Unified data warehouse - 'data-related roles can access whole data'
• pervasive automation - "Oh, I think we can do supervised learning on this task for automation"
• new roles such as Machine Learning Engineer

The process of AI transformation:

1. Execute pilot projects to gain momentum
2. Build an in-house AI team
3. Provide broad AI training
4. Develop own AI strategy
5. develop internal and external communication with various stakeholders
   For more information, check AI transformation playbook.

What Machine learning can and can not:

ML good at:

• Learning simple concept. Anything human can think in 1 second of thought.
• When we have lots of data available. ex) refund mail and refund datas

ML bad at:

• learning complicated concept with small amounts of data
• perform on new type of data. trained type and input type is different

For example,

• O : AI can detact where the car is in the photo
• X : AI cannot do knowing the intention of human hand gesture. -> hard to get enough data, need high accuracy
• O : AI can diagnose from 10,000 labeled X-ray images
• X : AI can not diagnose from only 10 labeled X-ray images with medical textbook explaining the disease