There's no question that AI is changing the way we work and live.

As reported in a study by McKinsey:

AI could potentially deliver additional economic output of around $13 trillion by 2030, boosting global GDP by about 1.2 percent a year.

It is really difficult to even conceive an industry that AI won't have a huge impact on.

The problem however, is that most CEOs of large companies, entrepreneurs, and employees don't understand how the technology really works. By the end of the article we're going to change that.

In this guide we will provide a complete overview of how to navigate the rise of AI, which will put you in a position to provide leadership on the topic to others.

We'll cover what's behind the all the buzzwords and how you can use AI for yourself, either in your personal or professional life.

This guide is based on Andrew Ng's AI for Everyone course, and is organized as follows:

  1. Artificial Narrow Intelligence (ANI) vs. Artificial General Intelligence (AGI)
  2. Important AI Terminology
  3. What is Machine Learning?
  4. What is Data?
  5. What is Deep Learning?
  6. What AI Can & Cannot Do
  7. How to Execute an AI Project
  8. How to Build AI Into Your Company
  9. Andrew Ng's AI Transformation Playbook
  10. Major Applications of Artificial Intelligence
  11. Understanding AI Techniques
  12. Conclusion: AI for Business

1. Artificial Narrow Intelligence (ANI) vs. Artificial General Intelligence (AGI)

As with any new technology there is a lot of hype surrounding AI. As Andrew Ng describes, one of the reasons for this is that AI is actually two separate ideas: ANI and AGI.

Artificial Narrow Intelligence (ANI)

Almost all of the progress we're seeing is coming from ANI, which refers to AI that does one thing such self driving cars, stock trading, and searching the web They are one trick ponies but when the AI gets it right, they can be incredibly valuable.

Artificial General Intelligence (AGI)

This refers to AI that can do anything a human can do, and most likely have superintelligence. We're seeing almost no progress in AGI at the moment.

Why it's important to separate these two is that people have falsely assumed that the advances in AI they see in the media are advances in AGI, instead of ANI, which can lead to a lot of irrational fear.

AGI will take many more technological breakthroughs to get to and may take decades to achieve this. In this article we're going to learn about artificial narrow intelligence.

One of the major technologies driving advances in AI are from machine learning, so let's discuss that now.

2. Important AI Terminology

Let's now familiarize ourselves with the most important concepts and terminology in AI so you can start thinking about how each one can be applied in your business. We'll get into a few of these topics in more detail, but here's a quick introduction of each.

Machine Learning

Machine learning is a subset of artificial intelligence based on the idea that systems can dynamically update their own parameters and learn from data, identify patterns and make decisions with minimal human intervention.

The term was coined by Arthur Samuel in 1959 as a:

A field of study that gives computers the ability to learn without being explicitly programmed.

A machine learning project will often result in software that can make the desired inference.

Data Science

Data science is process of analyzing a dataset in order to extract knowledge and insights from both structured and unstructured data.

Scientific methods are used to gain insights into our data, but the output will often be a PowerPoint presentation that summarizes conclusions for other teams in order for them to make better business decisions.

Deep Learning

Deep Learning is part of a broader family of machine learning and is based on learning data representations, as opposed to task-specific algorithms.

Deep learning makes use of neural network architectures, which are information processing patterns that are vaguely inspired by by our biological brains (but in reality completely different).

Deep Learning algorithms can be supervised, semi-supervised or unsupervised.

Neural Network

A neural network is not itself an algorithm, but rather a framework for many different machine learning algorithms that can work together and process complex data inputs.

As we can see from the image below, a neural network is really just what happens between our input data, and our desired output, and this is called a hidden layer.

A Deep Neural Network (DNN) is just a neural network that has multiple layers between the input and output layers.

Deep Learning and neural networks are pretty much used interchangeably today.

3. What is Machine Learning?

The rise of AI in the past few years has largely been driven by one thing: machine learning. Let's look at what it is so that you can start thinking about how you can apply it in your business.

Supervised Learning

The most common type of machine learning is called Supervised Learning.

Supervised Learning is the process of learning how to go from A to B, or mapping the input to output.

Here are a few examples to understand this:

  • Our input A could be an image of animal and we want to know the output B - what type of animal is it?
  • We could have an email as input A, and the output B is whether or not it is spam or not.
  • We could have an input A as English text and we want to output B the text in another language

One of the largest uses of supervised learning in industry is in online advertising, here's an overview of how it works:

  • The ad exchange receives input A, which is information about the ad and about you
  • The output B is whether or not you will click the ad or not.

The idea of supervised learning has been around for several decades, but has really taken off in the last few years. One of the main reasons for this is the explosion of big data. The other reason is due to the rise of neural networks and deep learning.

Here's how these two are connected:

  • Traditional AI systems saw that their performance did not continue to increase as you added more data, at a certain point there would be diminishing returns
  • With modern AI and neural networks, the performance generally keeps getting better as you add more data (to a certain point, but much longer than before)

For applications where having a highly accurate, high performance AI system is extremely important - let's say for self-driving cars - this made a huge difference, and much more valuable to companies and users.

The final piece in the puzzle is that these large neural networks require a huge amount of computing, and thanks to Moore's Law, cloud computing, and specialized processors like GPUs and TPUs (we'll get into these later) this computing power has now become accessible to companies of all sizes.

To sum it up - machine learning, and specifically supervised learning, has been the one of the most important ideas driving the advances in applied AI.

4. What is Data?

We've mentioned that data is extremely important in building effective AI solutions, so let's take a moment and answer the question: what is data?

Data can come in a variety of formats, but the most common is definitely a table of data (think excel spreadsheets), also referred to as a dataset. A table of data could contain information about stock prices, images of cats and dogs with their associated labels, etc.

How do we acquire data?

Manual Labelling

One way to acquire data is through manual labelling, for example we might collect a bunch of images of animals and manually label them with their species.

This can a very labor-intensive task since we might need hundreds of thousands of pictures to create a robust model, although there are few companies that offer data-labelling as a service such as thehive.ai and the CloudFactory.

Observing Behavior

Another way to acquire data is to observe the behaviors of users - for example, let's say we have an eCommerce store we can observe what visitors on our website are buying or not buying.In this case just by using our website users can generate meaningful data for us.

We can also observe the behavior of machines, for example to predict when they are going to stop working.

Downloading from Third-Party Websites

Thanks to the internet we can now download huge datasets for pretty much everything imaginable. A great example of this is Quandl, which has high quality data sources for Financial, Economic, and Alternative Data.

Here's a great article on 70 Free Data Sources from KDnuggets.

Common Misconceptions About Data

One of the common misconceptions about data is that you need to wait until you have a huge amount of data before applying machine learning algorithms to it. Instead you should generally start applying machine learning as soon as possible in order to get feedback about the data, and make iterative improvements from there.

Another misconception is that if you already have a good amount of data you can give it to an AI team and they can automatically make it valuable. While it is true that more data can often improve deep learning models with more data, this isn't always the case.

Data is Messy

Another thing to keep in mind is that if you're feeding an AI algorithm low-quality data, it can lead to it learning incorrect things.

An example of low-quality data are things like:

  • Incorrect labels
  • Missing values
  • Multiple types of data, for example having images, audio, and text - this is referred to unstructured data. Structured data, on the other hand, is data in the form of a table. Unstructured isn't necessarily worse than structured data - there are techniques for dealing with both types.

5. What is Deep Learning?

The terms neural network and deep learning are basically used interchangeably in artificial intelligence.

Coming back to our A input to B output formula, all a neural network really is the hidden layer between A and B. A deep neural network means there are more than 1 hidden layers between the input and output.

So what is the hidden layer actually doing to solve the problem of mapping input to output?

Each neuron is a set of inputs, a set of weights, and an activation function (which basically decides whether the neuron should fire or not).

The connections of the neurons are modeled as weights.

Each layer in a neural network consists of a number of neurons.

When you stack enough of these layers on top of each other, they can compute extremely complicated functions with a high degree of precision.

One of the great things about neural networks is that to train them all you have to do is give them the input A and desired output B, and it figures out the appropriate weights for each neuron by itself.

It turns out that if you provide a neural network enough data to train with it can do an incredibly good job of mapping input A to output B.

Any company today can make use of a few neural network or deep learning algorithms, but this doesn't automatically turn it into an AI-first company.

One aspect that makes a company great at AI is having a clear data acquisition strategy in place.

Many large tech companies today will keep unprofitable products in their lines, just for the sake of data acquisition. Having a keen understanding of where your data comes from, in what format, and where it is stored is key to a great AI company.

Many companies have taken efforts to bring their data into a single data warehouse to increase the chances their engineers can make sense of it.

AI-first companies also need to be great at spotting opportunities for automation in their business. They make use of machine learning engineers, data scientists, and subject-matter experts to discover new opportunities for growth and efficiency.

Of course, there is a systematic way that a company can accomplish this.

Here is the 5 Step AI Transformation Playbook from Andrew Ng:

  1. Execute pilot projects to gain momentum
  2. Build an in-house AI team
  3. Provide broad AI training
  4. Develop an AI strategy
  5. Develop internal and external communications

To sum it up - for a company to be great at AI means they have an architecture in place to make use of their data and allow machine learning algorithms to do what they do best.

6. What AI Can & Cannot Do

It's very common for most people to have overinflated expectations about what AI can do. This results in them asking engineers to do things that modern AI just hasn't cracked yet.

In this section we'll solve that by discussing exactly what AI can and can not do.

As Andrew Ng describes an imperfect rule of thumb for deciding if supervised learning can solve a task is:

If you can solve the task with 1 second of thought, we can probably now or soon automate it.

For example, in the case of self-driving car you can pretty much determine the location of another car in less than a second of thought. The input A is image data with another car, and the output B is the relative distance from your car.

If you're trying to figure out if an image is a cat or dog, you can figure it out in less than a second...you get the idea.

An example of something AI can probably not yet do is analyze a financial market and write a 20-page market analysis report on its findings, at least not that I know of. Actually I think I saw some like this last week...

The point to take away is that machine learning is an incredibly powerful tool, but it's also not magic and can't do everything (yet).

7. How to Execute an AI Project

Whether or not you want to play around with AI by yourself, or build AI into an entire corporation, this section will discuss how to actually go about executing an AI project.

Here's what we'll look at:

  • The Workflow of Artificial Intelligence
  • The Workflow of Data Science
  • How you choose an AI project to work on
  • How do we organize data and work with an AI team

Workflow of Artificial Intelligence

We can use machine learning algorithms to map input A to output B, but what are the steps we take to get there.

We get into this topic in much more detail in our What is Machine Learning article, but here's a summary:

The 7 Steps of Machine Learning:

  1. Gather the data
  2. Prepare the data for training
  3. Choose a model to use, each algorithm is suited for different types of data
  4. Train the model with our training data
  5. Test the model with our testing data
  6. Tune our models hyperparameters to improve performance
  7. Deploy the model for our desired prediction, or inference

After we deploy our model in the wild we then need to continuously collect new data, monitor how well it's doing, and update it based on new findings.

The Workflow of Data Science

Unlike a machine learning project who's goal is prediction, or inference, the objective of a data science is generally to discover a set of actionable insights from your data.

Data is transforming pretty much every existing job function, from manufacturing to marketing, and having an understanding our how it is transforming your job is crucial to success.

Here are a few examples of how tools like data science and machine learning can affect your job:

Sales & Marketing

If we're in marketing and our objective is to increase the conversion ratio of a sales funnel - we can use data science to better understand things like buyer demographics, comparing AB test results, etc.

We can use machine learning for the same sales funnel optimization task to prioritize which leads are in which stage of the funnel, so we know which ones should be prioritized.

Another example of machine learning in marketing is for product recommendations.

Manufacturing

Data science can be used to optimize manufacturing lines.

Machine learning can be used for things like the final inspection of the product. This would likely be a computer vision task to detect defects in your product line in real time, which would reduce labor costs and improve quality.

Regardless of the data science task, we would use these 3 steps to gain actionable insight from our data:

The 3 Steps of Data Science

  1. Collect the data - in our funnel example this would likely come from the ad platform you're using and your website data
  2. Analyze the data - for example using exploratory data analysis (EDA)
  3. Suggest actions/hypotheses based on the insights from analysis

How to Choose an AI Project

How do we choose an AI project that is worthy to pursue?

This is the framework we follow to brainstorm exciting AI projects. Since we now know that AI can't do everything, we need to make sure what we're pursuing is actually feasible. We also want to make sure that the project is valuable for your business.

So the goal is to pursue projects that are valuable and are actually possible in terms of the current technology.

AI experts will tend to have a better understanding of what AI can actually do, but then we need domain experts to figure out what is most valuable for a company.

For the team, we then want both types of people: AI experts and domain experts.

We use the following framework when brainstorming AI projects:

  • Focus on automating tasks instead of automating jobs - for example we may want to use machine learning to automatically route emails to the right department, but still have a customer service team for handling the email
  • Next we want to ask - what are the main drivers value and growth in your business?
  • Finally we ask - what are major paint points that you currently experience in your business?

After we have brainstormed several ideas about AI projects, we then need to commit to one. AI projects can can several months to execute so it's important to perform due diligence first.

How to Perform Due Diligence on an AI Project

The two things we want to do before starting a new project is:

  • Technical due diligence
  • Business due diligence

Technical due diligence is the process of making sure the AI project we have chosen is actually feasible.

  • An important question to ask in the step is how much data we are going to need, and how are we going to acquire this data?
  • Finally, we need to get an idea of engineering complexity and timeline - meaning how many people we are going to need working on it and for how long.

In terms of business due diligence, the first thing we check is how the project will actually drive value:

  • Many AI projects drive value through lowering costs as they improve efficiency.
  • AI projects can also increase revenue through things like conversion rate optimization, or personalized product recommendations.

Another question to ask in your due diligence is the age old question: build or buy?

If you don't have an in-house team of machine learning engineers and data scientists it can be quite costly to build that team just for one specific project, so in this case it would be better to outsource the project. Also there are plenty of things in AI that already have a high-level of industry standard, so it makes no sense to build these yourself.

Working with an AI Team

The first step before working with an AI team is specify your goals and your acceptance criteria.

There are certain projects where having performance that is worse than 99% accurate is just unacceptable.For example if your goal was to classify spam emails, this actually requires a very high degree of accuracy because if you incorrectly classified an email from your grandma as spam 1 out of 100 times, users are not going to be happy.

Commonly Used Technical Tools in AI

One of the reasons AI has taken off in the past few years is due to the huge contributions to the open-source community.

There are many open-source frameworks that AI teams use to build their systems, a few notable ones include:

  • TensorFlow
  • PyTorch
  • Keras
  • MXNet
  • Scikit-learn
  • CNTK
  • Caffe

Research breakthroughs are also published, most notably on Arxiv and developers share open source code repositories on GitHub.

In addition to these websites, machine learning teams also make use of cloud computing hardware, in particular:

  • CPU - or Central Processing Unit, this is the computer processor for a regular personal computer
  • GPU - or Graphics Processing Unit, this was historically used for processing pictures (for example in a video game). It turns out that this hardware is also well suited for building large neural networks, and has been another important factor in the advances in deep learning.
  • TPU - a Tensor Processing Unit is an AI accelerator developed by Google specifically for training large neural networks.

These computing resources are generally offered through cloud services like AWS and Google, but can also be done on-premise for larger enterprises.

You may have heard of Edge AI - which refers to the inference of a model happening on the specific device instead of having the data be sent to the cloud and back.

An example of why you would need this is in self-driving cars where you need the actions to be in near-real time, in which case you add a processor to the vehicle.

You can read more about Edge AI in our article: What is Edge AI? Value Propositions and Industry Use Cases.

8. How to Build AI Into Your Company

Now that you have a good understanding of what artificial intelligence actually is, and how to execute AI projects, let's look at how these projects fit in the context of a company.

Keep in mind that if you want to really get good at implementing AI in your company it will take 2-3 years to get enough experience with projects.

Building Complex AI Products

In a complex AI product there generally isn't just one machine learning algorithm that maps input A to output B. Instead the learning algorithms are often combined together and a part of a larger data structure and product.

Let's take a look at how a smart speaker like Siri or Alexa might process natural language commands:

  1. The first step is to listen for a trigger word, like 'Siri' or 'Alexa'
  2. The speaker needs to use speech recognition to process and transcribe the words that the user said, let's say it's 'Order me something from Amazon'
  3. The speaker then needs to recognize the intent from the speech, which is ordering a product. This is a process of mapping input A (natural language) to output B (intent of the command)
  4. Finally, the speaker needs to execute the command and order the desired product

This is a simple example of what we're doing in an AI pipeline.

To generalize this, the key steps in executing really any command on a smart speaker are:

  1. List for trigger word
  2. Speech recognition
  3. Intent recognition
  4. Run program to execute the command

To reinforce this point, let's look at an overview of what this AI pipeline might look like for a self-driving car.

Steps in the self-driving car AI pipeline:

  1. Detect Image, Radar, and Lidar data and combine this with existing GPS and map data
  2. Given the input data it needs to detect other cars, pedestrians, lanes, and traffic lights
  3. We then feed this into motion planning software, which plots appropriate paths for the car to take given the input data
  4. Finally we translate this into driving (steering, braking, and accelerating)

What roles are there in an AI team?

Large scale AI projects can sometimes require 100+ engineers working on it, so let's now review what a few of the important roles look like. Even if you're working with just a few people on a project, this will still be useful to get a sense of the type of jobs that an AI team would have.

As AI is advancing quickly, the exact job titles and responsibilities are not that well defined yet and can vary across companies.

Software Engineer

  • It's not uncommon for an AI team to be made up of 50%+ software engineers.

Machine Learning Engineer

  • These engineers are responsible for writing the software that executes the machine learning algorithm - either mapping an input to output (in the case of supervised learning), finding new structures/patterns in the data (with unsupervised learning), or maximizing an agents long-term expected reward (in reinforcement learning)

Machine Learning Researcher

  • The goal of this role is to advance the field of machine learning, either for public or private use.

Data Scientist

  • This role is not yet clearly defined, but one of the primary responsibilities is to examine data and provide actionable insights that are presented to their team/executives

Data Engineer

  • Due to the rise of big data, one of their primary responsibilities of a data engineer is organizing data. This means organizing data into an accessible, secure, and cost effective way. For many companies managing the volume of data is a huge task, so this is a highly sought after skill - for example in the case of self-driving cars they are collecting, analyzing, and acting on multiple movie-equivalent (gigabytes) worth of data every minute.

AI Product Manager

  • An AI product managers job is to figure out what is feasible to build and what is most valuable to a company, similar to a traditional product manager.

9. Andrew Ng's AI Transformation Playbook

Now that we've learnt what AI is, how to execute an AI project, and what typical job titles of an AI team are...let's look deeper at Andrew Ng's AI Transformation Playbook, which provides a roadmap for transforming your company with AI.

So what does it take to go from a great company to a great AI company?

Learning about this process is not only helpful to CEOs, but everyone in the company since AI will likely affect the entire company as a whole.

5 Steps of the AI Transformation Playbook

  1. Start with pilot projects to gain experience and momentum
  2. Build an in-house AI team
  3. Provide AI training to every level of the company, not just engineers
  4. Develop an AI strategy
  5. Develop an internal and external communications strategy for how you're using AI

Let's look at each in a bit more detail.

1. Pilot Projects

At this stage it's usually more important to execute small projects successfully than it is to tackle the biggest, most valuable problem in the company.

Ideally these projects are able to show traction and results within a 1 year timeframe. At this stage if you don't have an in-house AI team, outsourcing is still a great option and often advisable.

2. Building an In-House AI Team

An AI team should operate as its own separate, centralized business unit that reports directly to the CEO.

The talent from this team should then work with the team from every other business unit to support their work. The reason for this is generally that the other business units will be highly skilled at let's say advertising, or finances, but not be trained in AI yet, so having a separate, centralized team that can support these units makes sense.

3. Providing AI Training

As we've discussed in this article, a company needs AI talent in many more roles than just engineers, so making this training available for different teams is often smart to prioritize.

Usually this will start with just learning the basics of AI and an AI strategy so that they implement an optimal resource allocation. It's also important to continuously be providing the engineering team with training around the latest advances in the field.

4. Developing an AI Strategy

Developing an AI strategy is key to building and maintaining a competitive advantage in your specific industry.

CEOs often think that having an AI strategy should be step #1. This idea is often misguided and it is instead recommended to get your feet wet with a few smaller AI projects before building out an AI strategy.

It's also important to design an AI strategy that has a positive feedback loop, or as Andrew Ng calls it a "Virtuous Cycle of AI".

The Virtuous Cycle of AI is:

  • You start with a simple product and a small amount of data
  • You then get the product out to more users
  • This gives you more and more data
  • The product keeps getting better and better

Next you want to consider creating a strategic data acquisition strategy. An example if this is offering a free app or service that you can offer users that let's you collect and build a relevant dataset.

At this stage you may also consider setting up a unified data warehouse.

5. Develop an Internal & External Communications Strategy

This refers to things like:

  • Investor relations
  • Government relations
  • Consumer education
  • Talent recruitment
  • Internal communications

How to Take Your First Steps in AI

Now that we've looked at the 5 steps Andrew Ng suggests for transforming your company with AI, let's discuss how to take your first steps.

Here are a few different ways you can take your first steps in AI and start to build momentum:

  • Start brainstorming small pilot projects
  • Hire a few machine learning engineers and data scientists
  • Hire or appoint someone in your company as the AI leader
  • Have a discussion with your CEO or board about how much more valuable your company will be if you were great at AI

10. Major Applications of Artificial Intelligence

AI has been successfully applied to many areas including image and video data, language and speech data, and many others.

Let's take a closer look at a few areas that deep learning has excelled at.

Computer Vision

Computer vision generally refers to image classification and object recognition, either from image or video data.

Some of the major applications of computer vision include:

Facial Recognition

  • This area has had a lot of traction, and will certainly continue to grow in areas such as security.

Object Detection

  • For example detecting bikes on the road for self-driving cars.

Image segmentation

  • This takes object detection one step further, and tells us not only where the bike is, but can also segment each pixel of the bike compared to each pixel of a pedestrian, for example. The key point here is that it draws precise boundaries around each object.

Tracking

  • Computer vision can also track where the bike was 5 seconds ago, for example, which gives more information about speed, distance, etc. in order to figure out where the objects are going.

Natural Language Processing

Deep learning has also been making great strides in the field of Natural Langue Processing (NLP).

Here are a few of the major applications of NLP:

Text Classification

  • For example if we send an email to a customer support team, text classification can tell us if this email should go to the refund department, sales, etc.

Sentiment recognition

  • Text classification can also be taken a step further to sentiment recognition, where we can analyze the words used to determine the sentiment of the text.

Information Retrieval

  • The best known example of this is searching the web through a search engine. Companies will also use these internally for document retrieval.

Machine Translation

  • This is another major application of NLP - and refers to translating one language to another.

Speech Recognition

  • This refers to the problem of taking in an input of audio and determining what words are used.

Robotics

An obvious and major application of AI is in the field of robotics.

In robotics, the term perception refers to figuring out what is in the world around you based on the senses it has (cameras, radar, lidar).

  • Motion planning refers to finding the best possible path for the robot to follow.
  • Control refers to sending commands to the robot to follow a specific path.

11. Understanding AI Techniques

As mentioned earlier, Supervised Learning is currently the most common machine learning technique used today.

Let's now take a look at two more techniques called:

  • Unsupervised Learning
  • Reinforcement Learning

Unsupervised Learning

The best know example of unsupervised learning is called Clustering.

Here's an example of clustering...

Let's say you own a retail store that sells clothes. You can use data about your customers to figure out how many articles of clothing a customer buys and what the average price of each article of clothing.

Given data like this, a clustering algorithm can tell us that we have two groups in our data - customers that buy more inexpensive clothes buy more pieces, and customers that buy more expensive clothes buy fewer.

Clustering algorithms are commonly used in market segmentation and can help you discover insight about your customers.

The reason it is called Unsupervised Learning as opposed to Supervised is because with supervised learning we are providing the algorithm with labelled training data. In this case we have to provide it with the output B that we want.

With Unsupervised Learning we do not provide labelled data - we do not have to tell the algorithm exactly what to look for, instead we give it data (without any specific desired output labels) and we tell it to find something meaningful in the data.

The infamous example of Unsupervised Learning is Google Brain watching YouTube videos, and without being told what a cat is...it learned how to recognize cats.

One of the criticisms of Supervised Learning is that you need to provide it with so much labelled training data, which isn't always feasible at scale. This is also what makes Unsupervised Learning so valuable and a promising area of research.

Transfer Learning

Another important AI technique is called Transfer Learning.

Let's say that we have trained an algorithm to detect cars with 100,000 images, and then you want to teach it to recognize tractors...but own have 100 images.

Transfer Learning is the process of learning from task A (detecting cars) and using that knowledge to help learn task B (detecting tractors).

Reinforcement Learning

Another exciting field in AI is called Reinforcement Learning.

Let's say we want to create an AI that can fly a drone autonomously. This is not really possible for Supervised Learning because we can't always provide labelled training data for the optimal way to fly.

The problem that we're trying to solve here is somewhere in between Supervised and Unsupervised Learning.

In Reinforcement Learning, we have time-delayed labels that are sparse.

From these labels, which we can call rewards, we can learn to operate in this uncertain environment.

What we're doing with these rewards is basically telling the algorithm when it is performing well, and when it is doing poorly.

With this information the algorithm can learn how to maximize it's rewards over time.

Generative Adversarial Networks (GANs)

GANs are very good at creating new images from scratch.

Because of this, there is a a lot of work being done with GANs in the entertainment industry.

Understanding these algorithms does take some work, but is an extremely valuable skill leading to meaningful conversations with AI engineers.

12. Summary: AI for Business

We've now looked at:

  • Artificial Narrow Intelligence (ANI) vs. Artificial General Intelligence (AGI)
  • Important AI Terminology
  • What is Machine Learning?
  • What is Data?
  • What is Deep Learning?
  • What Can AI Do & Not Do
  • How to Execute an AI Project
  • How to Build AI Into Your Company
  • Andrew Ng's AI Transformation Playbook
  • Major Applications of Artificial Intelligence
  • Understanding AI Techniques

After understanding all of this, it's easy to see why AI can be such a superpower for teams of all sizes.

Whether you want to become a machine learning engineer, incorporate AI into your company, or be an AI thought leader in society you now have a solid foundation of some of the most important concepts.

To finish off this article it's important to remember that AI can't do everything, but it will certainly transform every industry in the coming years.