AI For Everyone -Andrew Ng - week 1

Week 1 What is AI?

Lesson 1 : Introduction

Lesson 2 Machine Learning 

In summary, supervised learning just learns input, output, or A to B mappings. On one hand, input, output A to B seems quite limiting.But when you find the right application scenario, this turns out to be incredibly valuable.

The most important idea in AI has been machine learning and specifically, supervised learning, which means A to B or input output mappings. What enables it to work really well is data. 

Lesson 3 What is the data?

 think through what is the data that is actually the most valuable

Germs of AI today is used primarily to generate unstructured data, such as text, images, and audio, rather than structured data. In contrast, supervised learning can work very well for both of these types of data, unstructured data and structured data. 

Lesson 4 AI terminology

• it's a piece of software that any time of day, any time of night, can automatically input A these properties of a house and output B. So if you have an AI system running serving dozens or hundreds of thousands of millions of users, that's usually a machine learning system.
• a data science project is a set of insights that can help you make business decisions, such as what type of house to build or whether to invest in renovation.  

Lesson 5 What makes an AI company?

• many great AI companies have preemptively invested in bringing the data together into single data warehouse to increase the odds that the teams can connect the dots.
• An Internet company is a company that does the thing that internet lets you do really well. 

• AI companies are very good at strategic data acquisition. This is why many of the large consumer tech companies may have three products that do not monetize and it allows them to acquire data that they can monetize elsewhere. Serve less strategy teams where we would deliberately launch products that do not make any money just for the sake of data acquisition.

This is the five-step AI transformation playbook that recommended to companies that want to become effective at using AI.

Step one is to execute pilot projects(试点项目) to gain momentum

just to a few small projects to get a better sense of what AI can or cannot do and get a better sense of what doing an AI project feels like.

Step two which is the building in house AI team and provide broad AI training

not just to the engineers but also to the managers, division leaders and executives and how they think about AI.

Step three have a better sense of what AI is and then is important for many companies to develop an AI strategy.

Finally, to align internal and external communications so that all your stakeholders from employees, customers and investors are aligned with how your company is navigating the rise of AI. 

Lesson 6 What machine learning can and cannot do

The main problem with this application, is that, the past history of a stock price is just not very predictive of the future stock price, which is why attempts to use machine learning this way haven't been successful. 

Future stock prices are so random that it's just hard for AI to predict it accurately. By the way, for completeness, I should say that predicting stock price based only on the historical price of the same stock seems to be impossible, but there are stock traders that sometimes find other inputs. For example, if they managed to legally obtain some web traffic or foot traffic data that helps them estimate what the company's sales were, then that, in combination with the historical price data, might make it possible for the algorithm to have some predictive power, however, these other inputs are typically complex or costly to acquire, and still can't overcome the intrinsically, somewhat random nature of the stock market.

Lesson 7 More examples of what machine learning can and cannot do

 if you were to try to build a system to learn the A to B mapping, where the input A is a short video of our human gesturing at your car, and the output B is, what's the intention or what does this person want, that today is very difficult to do. Part of the problem is that the number of ways people gesture at you is very, very large. Imagine all the hand gestures someone could conceivably use asking you to slow down or go, or stop. The number of ways that people could gesture at you is just very, very large. So, it's difficult to collect enough data from enough thousands or tens of thousands of different people gesturing at you, and all of these differentways to capture the richness of human gestures. 

Lesson 7 Non-technical explanation of deep learning (Part 1, optional)​​​​​​​

It's a group of artificial neurons, each of which computes a relative simple function. But when you stack enough of them together, like lego breaks, they can compute incredibly complicatedfunctions that give you very accurate mappings from the input A to the output B. 

Lesson 8 Non-technical explanation of deep learning (Part 2, optional)

Reference

AI For Everyone - What is AI? - Week 1 | Coursera