Artificial Intelligence

43 minute read

Published:

This lesson is from AI For Everyone

Introduction

  • According to a study by McKinsey Global Institute, AI is estimated to create an additional 13 trillion US dollars of value annually by the year 2030.
  • Even though AI is already creating tremendous amounts of value into software industry, a lot of the value to be created in a future lies outside the software industry.
  • In sectors such as retail, travel, transportation, automotive, materials, manufacturing and so on.
    • AI will have a huge impact on in the next several years.
    • Hairdressing industry because we know how to use AI robotics to automate hairdressing.
  • There is a lot of excitement but also a lot of unnecessary hype about AI.
    • One of the reasons for this is because AI is actually two separate ideas.
    • Almost all the progress we are seeing in the AI today is artificial narrow intelligence.
    • These are AIs that do one thing such as a smart speaker or a self-driving car or AI to do web search or AI applications in farming or in a factory.
    • These types of AI are one trick ponies but when you find the appropriate trick, this can be incredibly valuable.
  • Unfortunately, AI also refers to a second concept of AGI or artificial general intelligence.
    • That is the goal to build AI.
    • They can do anything a human can do or maybe even be superintelligence and do even more things than any human can.
    • There is tons of progress in ANI, artificial narrow intelligence and almost no progress to what AGI or artificial general intelligence.
  • Both of these are worthy goals and unfortunately the rapid progress in ANI which is incredibly valuable, that has caused people to conclude that there’s a lot of progress in AI, which is true.
  • But that has caused people to falsely think that there might be a lot of progress in AGI as well which is leading to some irrational fears about evil clever robots coming over to take over humanity anytime now.
    • I think AGI is an exciting goal for researchers to work on, but it’ll take most for technological breakthroughs before we get there and it may be decades or hundreds of years or even thousands of years away.
    • Given how far away AGI is, I think there is no need to unduly worry about it.
  • You will learn what ANI can do and how to apply them to your problems.

Machine Learning

  • The rise of AI has been largely driven by one tool in AI called machine learning.
  • The most commonly used type of machine learning is a type of AI that learns A to B, or input to output mappings.
  • This is called supervised learning.
  • Let’s see some examples.
    • If the input A is an email and the output B one is email spam or not, zero one.
      • Then this is the core piece of AI used to build a spam filter.
    • If the input is an audio clip, and the AI’s job is to output the text transcript,
      • then this is speech recognition.
    • If you want to input English and have it output a different language, Chinese, Spanish, something else,
      • then this is machine translation.
  • the most lucrative form of supervised learning, of this type of machine learning maybe be online advertising
    • where all the large online ad platforms have a piece of AI that inputs some information about an ad, and some information about you, and tries to figure out,
      • will you click on this ad or not?
      • By showing you the ads you’re most likely to click on, this turns out to be very lucrative.
      • Maybe not the most inspiring application, but certainly having a huge economic impact today.
  • if you want to build a self-driving car, one of the key pieces of AI is in the
    • AI that takes as input an image, and some information from their radar, or from other sensors, and
      • output the position of other cars, so your self-driving car can avoid the other cars.
  • Manufacturing.
    • where you take as input a picture of something you’ve just manufactured, such as a picture of a cell phone coming off the assembly line.
    • This is a picture of a phone, not a picture taken by a phone, and you want to output, is there a scratch, or is there a dent, or some other defects on this thing you’ve just manufactured?
    • And this is visual inspection which is helping manufacturers to reduce or prevent defects in the things that they’re making.
  • This set of AI called supervised learning, just learns input to output, or A to B mappings.
  • On one hand, input to output, A to B it seems quite limiting. But when you find a right application scenario, this can be incredibly valuable.
  • Now, the idea of supervised learning has been around for many decades. But it’s really taken off in the last few years.
  • AI has really taken off recently due to the rise of neural networks and deep learning.
  • But with modern AI, with neural networks and deep learning, what we saw was that, if you train a small neural network, then the performance looks like this, where as you feed them more data, performance keeps getting better for much longer.
  • If you train a even slightly larger neural network, say medium-sized neural net, then the performance may look like that. If you train a very large neural network, then the performance just keeps on getting better and better.
  • For applications like speech recognition, online advertising, building self-driving car, where having a high-performance, highly accurate, say speech recognition system is important, enable these AI systems get much better, and make speech recognition products much more acceptable to users, much more valuable to companies and to users.
  • The most important idea in AI has been machine learning, has basically supervised learning, which means A to B, or input to output mappings.
  • What enables it to work really well is data.

Data

  • Data is really important for building AI systems.

  • House Price
    • first column: size of the house, say in square feet or square meters, and the
    • second column: price of the house.
  • Build a AI system or Machine Learning system to help you set prices for houses or figure out if a house is priced appropriately,
    • you might decide that the size of the house is A and the price of the house is B, and have an AI system learn this input to output or A to B mapping.
  • Collect data on the number of bedrooms of this house.
    • A can be both of these first two columns, and B can be just the price of the house.
  • So, given that table of data, given the dataset, it’s actually up to you, up to your business use case to decide what is A and what is B.
  • Data is often unique to your business, and this is an example of a dataset that a rural state agency might have that they tried to help price houses.
  • It’s up to you to decide what is A and what is B, and how to choose these definitions of A and B to make it valuable for your business.
  • As another example, if you have a certain budget and you want to decide what is the size of house you can afford, then
    • you might decide that the input A is how much does someone spend and
    • B is just the size of the house in square feet, and
    • that would be a totally different choice of A and B that tells you, given a certain budget, what’s the size of the house you should be maybe looking at.
  • Let’s say that you want to build a AI system to recognize cats in pictures.
    • dataset where the input A is a set of different images and the
    • output B are labels that says, “First picture is a cat, that’s not a cat.
    • That’s a cat, that’s not a cat” and have an AI input a picture A and output B is it the cats or not so you can tag all the cat pictures on your photo feed or your mobile app.
  • So, data is important. But how do you get data? How do you acquire data?
  • Well, one way to get data is manual labeling.
  • For example, you might collect a set of pictures, and then you might either yourself or have someone else go through these pictures and label each of them.
  • So, the first one is a cat, second one is not a cat, third one is a cat, fourth one is not a cat.
  • By manually labeling each of these images, you now have a dataset for building a cat detector.
  • You might need hundreds of thousands of pictures but manual labeling is a tried and true way of getting a dataset where you have both A and B.
  • Another way to get a dataset is from observing user behaviors or other types of behaviors.
  • So, for example, let’s say you run a website that sells things online.
  • So, an e-commerce or an electronic commerce website where you offer things to users at different prices, and you can just observe if they buy your product or not.
  • So, just through the act of either buying or not buying your product, you may be able to collected a data set like this,
    • where you can store the user ID, the time the user visited your website, the price you offer the product to the users as well as
    • whether or not they purchased it.
  • So, just by using your website, users can generate this data from you.
  • This was an example of observing user behaviors.
  • We can also observe behaviors of other things such as machines.
  • If you run a large machine in a factory and you want to predict if a machine is about to fail or have a fault, then just by observing the behavior of a machine, you can then record a dataset like this.
  • There’s a machine ID, there’s a temperature of the machine, there’s a pressure within the machine, and then did the machine fail or not.
  • If your application is prevent the maintenance, say you want to figure out if a machine is about to fail, then you could for example, choose this as the input A and choose that as the output B to try to figure out if a machine is about to fail in which case you might do preventative maintenance on the machine.
  • The third and very common way of acquiring data is to download it from a website or to get it from a partner.
  • Thanks to the open internet, there’s just so many, there’s as that you can download for free, ranging from computer vision or image datasets, to self-driving car datasets, to speech recognition datasets, to medical imaging data sets to many many more.
  • So, if your application needs a type of data, you just download off the web keeping in mind licensing and copyright, then that could be a great way to get started on the application.
  • Finally, if you’re working with a partner, say you’re working with a factory, then they may already have collected a big dataset, machines, and temperatures, and pressure into the machines fail not that they could give to you.
  • Data is important, but there’s also little bit over-hyped and sometimes misused.
    • Waiting to collecting Data
      • give me three years to build up my IT team, we’re collecting so much data. Then after three years, I’ll have this perfect dataset, and then we’ll do AI then.”
      • once you’ve started collecting some data, go ahead and start showing it or feeding it to an AI team.
      • Because often, the AI team can give feedback to your IT team on what types of data to collect and what types of IT infrastructure to keep on building.
      • For example, maybe an AI team can look at your factory data and say, “Hey. You know what? If you can collect data from this big manufacturing machine, not just once every ten minutes, but instead once every one minute, then we could do a much better job building a preventative maintenance systems for you.”
      • So, there’s often this interplay of this back and forth between IT and AI teams, and advise is usually try to get feedback from AI earlier, because it can help you guide the development of your IT infrastructure.
    • misuse of data.
      • More data is usually better than less data, but I wouldn’t take it for granted that just because you have many terabytes or gigabytes of data, that an AI team can actually make that valuable.
      • So, don’t throw data in a AI team and assume it will be valuable.
      • Don’t over-invest in just acquiring data for the sake of data until unless you’re also getting an AI team to take a look at it. Because, they can help guide you to think through what is the data that is actually the most valuable.
    • data is messy.
      • You may have heard the phrase garbage in garbage out, and if you have bad data, then the AI will learn inaccurate things.
      • Here are some examples of data problems. Let’s say you have this data sets of size of houses, number of bedrooms, and the price. You can have incorrect labels or just incorrect data.
        • this house is probably not going to sell for $0.1 just for one dollar.
        • data can also have missing values such as we have here a whole bunch of unknown values.
        • So, your AI team will need to figure out how to clean up the data or how to deal with these incorrect labels and all missing values.
      • There are also multiple types of data.
        • For example, sometimes you hear about images, audio, and text.
        • These are types of data that humans find it very easy to interpret.
        • There’s a term for this. This is called unstructured data, and there’s a certain types of AI techniques that could work with images to recognize cats or audios to recognize speech or texts, or understand that email is spam.
        • Then, there are also datasets - structured data.
          • That basically means data that lives in a giant spreadsheet, and the techniques for dealing with unstructured data are little bit different than the techniques for dealing with structured data.
          • But AI techniques can work very well for both of these types of data, unstructured data and structured data.

Terminology

  • machine learning or data science or neural networks or deep learning.

  • Let’s say you have a housing dataset like this with the size of the house, number of bedrooms, number of bathrooms, whether the house is newly renovated as was the price.

  • If you want to build a mobile app to help people price houses, so input A, and output B.

    • Machine-learning system, and particular would be one of those machine learning systems that learns inputs to outputs, or A to B mappings.
    • So, machine learning often results in a running AI system.
    • So, it’s a piece of software that anytime of day, anytime of night you can automatically input A these properties of 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.
    • In contrast, here’s something else you might want to do, which is to have a team analyze your dataset in order to gain insights.
      • So, a team might come up with a conclusion like,
      • “Hey, did you know if you have two houses of a similar size, they’ve a similar square footage, if the house has three bedrooms, then they cost a lot more than the house of two bedrooms, even if the square for this is the same.” Or,
      • “Did you know that newly renovated homes have a 15% premium, and this can help you make decisions such as, given a similar square footage, do you want to build a two bedroom or three bedroom size in order to maximize value? “ Or,
      • “Is it worth an investment to renovate a home in the hope that the renovation increases the price you can sell a house for?”
    • So, these would be examples of data science projects, where the output of 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.
    • The boundaries between these two terms, machine learning and data science are actually little bit buzzy, and these terms are not used consistently even in industry today.
    • But what I’m giving here is maybe the most commonly used definitions of these terms, but you will not find universal adherence to these definitions.
    • To formalize these two notions a bit more, machine learning is the field of study that gives computers the ability to learn without being explicitly programmed.
    • This is a definition by Arthur Samuel many decades ago. Arthur Samuel was one of the pioneers of machine learning, who was famous for building a checkers playing program. They could play checkers, even better than he himself, the inventor could play the game.
  • So, a machine learning project will often results in a piece of software that runs, that outputs B given A.

  • In contrast, data science is the size of extracting knowledge and insights from data.

    • So, the output of a data science project is often a slide deck, the PowerPoint presentation that summarizes conclusions for executives to take business actions or that summarizes conclusions for a product team to decide how to improve a website.
  • Let me give an example of machine learning versus data science in the online advertising industry.

  • Today, to launch our platforms, all have a piece of AI that quickly tells them what’s the ad you are most likely to click on. So, that’s a machine learning system.

  • This turns out to be incredibly lucrative AI system to inputs enrich about you and about the ad and outputs where you click on this or not. These systems are running 24-7. These are machine learning systems that drive our gravity for these companies, such as a piece of software that runs.

  • If analyzing data tells you, for example, that the travel industry is not buying a lot of ads,

    • but if you send more salespeople to sell ads to travel companies,
    • you could convince them to use more advertising,
    • then that would be an example of a data science project and the data science conclusion the results and the executives deciding to ask a sales team to spend more time reaching out to the travel industry.
  • So, even in one company, you may have different machine learning and data science projects, both of which can be incredibly valuable.

  • You have also heard of deep learning.

  • So, what is deep learning?

  • Let’s say you want to predict housing prices, you want to price houses. So, you will have an input that tells you the size of the house, number of bedrooms, number of bathrooms and whether it’s newly renovated.

  • One of the most effective ways to price houses, given this input A would be to feed it to NN here in order to have it output the price.

    • This big thing in the middle is called a neural network, and sometimes we also called an artificial neural network.
    • That’s to distinguish it from the neural network that is in your brain.
  • So, the human brain is made up of neurons.

  • So, when we say artificial neural network, that’s just to emphasize that this is not the biological brain, but this is a piece of software.

  • What a neural network does, or an artificial neural network does is takes this input A, which is all of these four things, and then output B, which is the estimated price of the house.

  • But all of human cognition is made up of neurons in your brain passing electrical impulses, passing little messages each other.

  • When we draw a picture of an artificial neural network, there’s a very loose analogy to the brain.

  • These little circles are called artificial neurons, or just neurons for short.

  • That also passes neurons to each other.

  • This big artificial neural network is just a big mathematical equation that tells it given the inputs A, how do you compute the price B.

  • Key takeaways are that a neural network is a very effective technique for learning A to B or input-output mappings.

  • Today, the terms neural network and deep learning are used almost interchangeably, they mean essentially the same thing.

  • Many decades ago, this type of software was called a neural network.

  • But in recent years, we found that deep learning was just a much better sounding brand, and so that for better or worse is a term that’s been taken off recently.

  • So, what do neural networks or artificial neural networks have to do with the brain?

    • It turns out almost nothing.
    • Neural networks were originally inspired by the brain, but the details of how they work are almost completely unrelated to how biological brains work.
  • Unsupervised learning, reinforcement learning

  • You don’t need to know what all of these other terms mean, but these are just other tools to getting AI systems to make computers act intelligently.

  • The part of machine learning that’s most important these days is neural networks or deep learning, which is a very powerful set of tools for carrying out supervised learning or A to B mappings as well as some other things.

  • But there are also other machine learning tools that are not just deep learning tools.

  • So, how does data science fit into this picture? There is inconsistency in how the terminology is used.

  • Some people will tell you data science is a subset of AI.

  • Some people will tell you AI is a subset of data science.

  • Data science is maybe a cross-cutting subset of all of these tools that uses many tools from AI machine learning and deep learning, but has some other separate tools as well that solves a very set of important problems in driving business insights.

AI Company

  • What makes a company good at AI? Perhaps even more importantly, what will it take for your company to become great at using AI?
  • An Internet company
    • is a company that does the thing that internet let you do really well.
    • For example, we engage and pervasive AB testing.
    • Meaning we routinely threw up two different versions of website and see which one works better because we can. So, we learn much faster.
      • Whereas in a traditional shopping mall, very difficult to have two shopping malls in two parallel universes and you can only maybe change things around every quarter or every six months.
      • Internet company is since a very short iteration times. You can ship a new product every week or maybe even every day because you can whereas a shopping mall can be redesigned and we are protected only every several months.
      • Internet companies also tend to push decision-making down from the CEO to the engineers and to other specialized rules such that the product managers.
      • This is in contrast to a traditional shopping mall. We can maybe have the CEO just decide all the key decisions and then just everyone does what the CEO says.
      • It turns out that traditional model doesn’t work in the internet era because only the engineers and other specialized roles like product managers know enough about the technology and the product and the users to make great decisions.
    • AI era?
      • I think that today, you can take any company and have it use a few neural networks or few deep learning algorithms.
      • That by itself does not turn the accompany into an AI company.
      • Instead, what makes a great AI company, sometimes an AI first company is, are you doing the things that AI lets you do really well?
      • For example, AI companies are very good at strategic data acquisition.
      • This is why many of the large consumer tech companies may have 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.
        • Thinking through how to get data is a key part of the great AI companies.
        • AI companies sends a unified data warehouses.
        • If you have 50 different databases or 50 different data warehouses under the control of 50 different Vice-Presidents, then there will be impossible for an engineer to get the data into one place so that they can connect the dots and spot the patterns.
        • So, 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.
          • Subject of course to privacy guarantees and also to data regulations such as GDPR in Europe.
        • AI companies are very good at spotting automation opportunities.
        • Let’s insert the supervised learning algorithm and have a mapping here so that we don’t have to have people do these tasks instead we can automate It.
        • AI companies also have many new roles such as the MLE or Machine Learning Engineer and new ways of dividing up tasks among different members of a team.
        • So, for a company to become good at AI means, architecting for company to do the things that AI makes it possible to do really well.
        • Now, for a company that become good at AI does require a process.
        • In fact, 10 years ago, Google and Baidu as well as companies like Facebook and Microsoft were not great AI companies the way they are today.
        • So, how can a company become good at AI?
        • It turns out that becoming good at AI is not a mysterious magical process.
        • Instead there is a systematic process through which many companies, almost any big company can become good at AI.
        • This is the five-step AI transformation playbook that I recommend to companies that want to become effective at using AI.
          • Step one is to execute pilot projects to gain momentum.
            • So, 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.
            • This you could do in house or you can also do with an outsource team.
          • Step two which is the building in house AI team
            • provide broad AI training, not just to the engineers but also to the managers, division leaders and executives and how they think about AI.
          • After doing this or as you’re doing this, you 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 aligns with how your company is navigating the rise of AI.
          • AI has created tremendous value in the software industry and will continue to do so.
          • It will also create tremendous value outside the software industry.
          • If you can help your company become good at AI, I hope you can play a leading role in creating a lot of this value.

What machine learning can and cannot do

  • Do technical diligence on the project to make sure that it is feasible. This means:
    • looking at the data, look at the input, and output A and B, and just thinking through if this is something AI can really do.
  • One of the challenges is that the media, as well as the academic literature, tends to only report on positive results or success stories using AI, and we see a string of success stories and no failure stories, people sometimes think AI can do everything.
  • Previously, you saw this list of AI applications from spam filtering to speech recognition, to machine translation, and so on.
  • One imperfect rule of thumb you can use to decide what supervised learning may or may not be able to do is that,
    • pretty much anything you could do with a second of thought,
    • we can probably now or soon automate using supervised learning, using this input-output mapping.
  • So for example, in order to determine the position of other cars, that’s something that you can do with less than a second.
  • In order to tell if a phone is scratched, you can look at it and you can tell in less than a second.
  • In order to understand or at least transcribe what was said, it doesn’t take that many seconds of thought.
  • While this is an imperfect rule of thumb, it maybe gives you a way to quickly think of some examples of tasks that AI systems can do.
  • Whereas in contrast, something that AI today cannot do would be:
    • to analyze a market and write a 50 page report, a human cannot write a 50 page mark of analysis report in a second, and it’s very difficult
  • I don’t think any team in the world today knows how to get an AI system to do market research and run an extended market report either.
  • Let’s take a look at a specific example, relating to customer support automation.
  • Let’s see a random website that sells things, so an e-commerce company, and you have a customer support division that gets an email like this, “The toy arrived two days late, so I wasn’t able to give it to my niece for her birthday. Can I return it?”
  • If what you want is an AI system that looks at this and decides this is a refund request, so let me route it to my refund department, then I will say, you have a good chance of building an AI system to do that.
  • The AI system would take as input, the customer text, what the customer emails you, and it would output, is this a refund requests or is this a shipping problem, or is it the other request, in order to route this email to the most appropriate parts of your customer support center.
  • So, the input A is the text and the output B is one of these three outcomes, is it a refund or a shipping problem, or shipping query, or is it a different requests.
  • So, this is something that AI today can do. Here’s something that AI today cannot do which is if you want the AI to input an email and automatically generate, it responds like, “Oh, sorry to hear that. I hope you’re niece had a good birthday. Yes, we can help with, and so on.”
  • So, for an AI to output a complicated piece of text like this today is very difficult by today’s standards of AI and in fact to even empathize about the birthday of your niece, that is very difficult to do for every single possible type of email you might receive.
  • Now, what would happen if you were to use a machine learning tool like a deep learning algorithm to try to do this anyway.
  • So, let’s say you tried to get an AI system to input the user’s email, and output a two to the three paragraph, empathetic and appropriate response.
  • Let’s say that you have a modest-sized dataset like a 1,000 examples of user emails and appropriate responses.
  • It turns out if you run an AI system on this type of data, on a small dataset like 1,000 examples, this may be the performance you get,
    • which is if a user emails, “My box was damaged,” they’ll say, “Thank you for your email,” and it says, “Where do I write a review?”, “Thank you email.” “What’s the return policy?”, “Thank you for your email.”
  • But the problem with building this type of AI is that with just a 1,000 examples, there’s just not enough data for an AI system to learn how to write to the three paragraph, appropriate and empathetic responses.
  • So, you may end up just generating the same very simple response like, “Thank you for your email,” no matter what the customer is sending you.
  • Another thing that could go wrong, another way for an AI system to fail is if it generates gibberish such as: “When is my box arriving,” and it says, “Thank, yes, now your,” gibberish.
    • This is a hard enough problem that even with 10,000 or a 100,000 email examples, I don’t know if that would be enough data for an AI system to do this well.
  • The rules for what AI can and cannot do are not hardened first and usually end up having to ask engineering teams to sometimes spend a few weeks doing deep technical diligence to decide if a project is feasible.
  • But to hone your intuitions to help you quickly filter feasible or not feasible projects, here are a couple of other rules of thumb about what makes a machine learning problem easier or more likely to be feasible.
  • One, learning a simple concept is more likely to be feasible.
  • Well, what does a simple concept mean?
  • There’s no formal definition of that but it is something that takes you less than a second of mental thought or a very small number of seconds of mental thought to come up with a conclusion then that would lean to whether it being a simple concept.
  • So, you’re looking outside the window of a self-driving car to spot the other cars that would be a relatively simple concept.
  • Whereas how to write an empathetic response, so a complicated user complaints, that would be less of a simple concept.
  • Second, a machine learning problem is more likely to be feasible if you have lots of data available.
  • Here, our data means both the input A and the output B, that you want the AI system to have in your A to B, input to output mapping.
  • So for example, in the customer support application, the input A would be examples of emails from customers and B could be labeling each of these customer emails as to whether it’s a refund requests or a shipping query, or some other problem, one of three outcomes.
  • Then if you have thousands of emails with both A and B, then the odds of you building a machine learning system to do that would be pretty good.
  • AI is the new electricity and it’s transforming every industry, but it’s also not magic and it can’t do everything under the sun.

More examples of what machine learning can and cannot do

  • Let’s say you’re building a self-driving car
    • here’s something that AI can do pretty well, which is to take a picture of what’s in front of your car and maybe just using a camera, maybe using other senses as well such as radar or lidar.
    • Then to figure out, what is the position, or where are the other cars.
    • So, this would be an AI where the input A, is a picture of what’s in front of your car, or maybe both a picture as well as radar and other sensor readings.
    • The output B is, where are the other cars?
    • Today, the self-driving car industry has figured out how to collect enough data and has pretty good algorithms for doing this reasonably well.
    • So, that’s what the AI today can do.
  • Here’s an example of something that today’s AI cannot do, or at least would be very difficult using today’s AI, which is to
    • input a picture and output the intention of whatever the human is trying to gesture at your car.
    • a construction worker holding out a hand to ask your car to stop.
    • a bicyclist raising the left-hand to indicate that they want to turn left.
    • So, 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 different ways to capture the richness of human gestures.
    • So, learning from a video to what this person wants, it’s actually a somewhat complicated concept.
    • In fact, even people have a hard time figuring out sometimes what someone waving at your car wants.
    • Then second, because this is a safety critical application, you would want an AI that is extremely accurate in terms of figuring out, does a construction worker want you to stop, or does he or she wants you to go?
    • And that makes it harder for an AI system as well.
    • So, today if you collect just say, 10,000 pictures of other cars, many teams would build an AI system that at least has a basic capability at detecting other cars.
    • In contrast, even if you collect pictures or videos of 10,000 people, it’s quite hard to track down 10,000 people waving at your car.
    • Even with that data set, I think it’s quite hard today to build an AI system to recognize humans intentions from their gestures at the very high level of accuracy needed in order to drive safely around these people.
    • So, that’s why today, many self-driving car teams have some components for detecting other cars, and they do rely on that technology to drive safely.
    • But very few self-driving car teams are trying to count on the AI system to recognize a huge diversity of human gestures and counting just on that to drive safely around people.
  • Say you want to build an AI system to look at X-ray images and diagnose pneumonia.
    • So, all of these are chest X-rays.
    • So, the input A could be the X-ray image and the output B can be the diagnosis.
    • Does this patient have pneumonia or not?
      • So, that’s something that AI can do.
    • Something that AI cannot do would be to diagnose pneumonia from 10 images of a medical textbook chapter explaining pneumonia.
    • A human can look at a small set of images, maybe just a few dozen images, and reads a few paragraphs from medical textbook and start to get a sense.
    • But actually don’t know, given a medical textbook, what is A and what is B? Or
    • how to really pose this as an AI problems like know how to write a piece of software to solve, if all you have is just 10 images and a few paragraphs of text that explain what pneumonia in a chest X-ray looks like.
    • Whereas a young medical doctor might learn quite well reading a medical textbook at just looking at maybe dozens of images.
    • In contrast, an AI system isn’t really able to do that today.
  • To summarize, here are some of the strengths and weaknesses of machine learning.
    • Machine learning tends to work well when you’re trying to learn a simple concept, such as something that you could do with less than a second of mental thought, and when there’s lots of data available.
    • Machine learning tends to work poorly when you’re trying to learn a complex concept from small amounts of data.
    • A second under-appreciated weakness of AI is that it tends to do poorly when it’s asked to perform on new types of data that’s different than the data it has seen in your data set.
  • Say you built a supervised learning system that uses A to B to learn to diagnose pneumonia from images like these.
    • These are well pretty high quality chest X-ray images.
    • But now, let’s say you take this AI system and apply it at a different hospital or different medical center, where maybe the X-ray technician somehow strangely had the patients always lie at an angle or sometimes there are other defects.
    • This is one of the things that AI is actually much weaker than humans.

Deep Learning

  • The terms deep learning and neural network are used almost interchangeably in AI.

  • Let’s say you run a website that sells t-shirts.

    • And you want to know, based on how you price the t-shirts, how many units you expect to sell, how many t-shirts you expect to sell.

    • You might then create a dataset like this, where the higher the price of the t-shirt, the lower the demand.

  • Face Recognition

    • Say you want to build a system to recognize people from pictures, how can a piece of software look at this picture and figure out the identity of the person in it?
    • If the resolution of this picture is 1000 pixels by 1000 pixels, then that’s a million pixels.
    • So, if it were a black and white or grayscale image, this neural network was take as input a million numbers corresponding to the brightness of all one million pixels in this image or
      • if was a color image it would take as input three million numbers corresponding to the red, green, and blue values of each of these one million pixels in this image.