It has been 3 years that I dove headfirst into the Artificial Intelligence universe. I joined the next startup, Nauto. A few weeks earlier Nauto released a groundbreaking new device. An in-vehicle camera that uses Computer Vision to detect if drivers are paying attention to the road.
An incredibly complex problem. And even though I had worked on Machine Learning enabled products before, I was up for a completely new challenge. It has been 3 years now and it is time to update the article I wrote back then.
3 years is a long time in such a nascent field of technology that moves rapidly. A lot of new applications leveraging AI came out. But the underlying principles and technologies haven't changed that much in 3 years. One thing is still certain. Artificial Intelligence will be a leading factor in technology for the years to come. New industries will be born out of it, existing ones will change completely.
If you have some years in your career ahead of you, AI will be right there with you. AI is already applied in a wide range of products to solve problems for companies. Internal and external (for customers) applications are getting tested and applied. As companies learn from each other and see the benefits, it will spill over from the early adopters to all.
If you start now to gain AI knowledge and understand the basics, it will, without a doubt, help you in your career.
The notes below are a personal summary. I am not a scientist or a math genius. I started with almost no experience working on AI/ML related products. If you are an expert, this might be too simplified. For me, it is exactly what I was looking for when I started. The goal for Product Managers is to understand what AI can do for their product and their customers. How can it enable them to live better lives? How can it save them time? Once we understand that, we can start thinking about what is the next cool thing we can build. And we will find new opportunities and features to work on. Or come up with ideas for completely new companies.
Decoding buzzwords
We hear a lot of words in context with AI. Deep, Neural, Convolutional... It is very confusing if you don’t know where these words belong. So I want to provide an overview below. The infographic shows you the categories and gives a quick description of what it is.
.png)
No decisions without data.
Now that we have a good understanding of the categories, let's dive into a fundamental concept.
Once we are clear on what our customer's problem is we want to solve, we can think about it this way. Let's say it is a decision problem. Like: What is the right toy for my 1-year-old toddler?The process is very similar to what you would do when making a decision. You gather information. This information will help you make a decision. And that is exactly what AI needs to do. We take this information and feed it to the AI as data. And we mimic in some ways how we humans learn and make decisions.
Now AI can do several things. We can train it to improve in that decision-making process. You keep feeding information and rate the outcome when the AI suggests a decision. This is the process of the AI to learn about right or wrong decisions. Cognitive capabilities. Deep Learning.
But you could also say, I want to base the decision on statistical data. Like, what others have bought and rated highly. Maybe others that match my persona and the persona of my toddler.
When AI makes decisions based on statistics it is statistical learning. Same as applied in Natural Language Processing.
Both forms I described are a category of Machine Learning. In a lot of cases, we use Artificial Intelligence and Machine Learning interchangeably. A system that we want to learn, like a car, we equip it with sensors. This way it gets new data analyzes it and then provides decisions. The more data points we can collect and use, the higher the chances that the accuracy (precision and recall) of our AI is higher.
Think about the toddler suggestions example above. This is not very impressive as it happens all the time on Amazon already. But think about it this way. When a never before seen customer joins they start to interact with the platform. The AI starts to build a persona and look for similarities. After seeing the person only for a few clicks it will already know products with a very high probability of relevance. That’s powerful.
How can we equip AI to collect data?
As you can see, data becomes gold, a valuable commodity. And so does the ability to collect data. Companies that understand this move into the direction to collect data. Companies that are leading their industry are doing it for many years. Sometimes to controversial extend and regulators are rightfully catching up.
We build cameras and databases of images to give our AI image data about the world. A very famous organized database is ImageNet. It is a database that includes 14 Million images categorized in subsets.
If you would want to go and teach a computer to recognize a banana, you could use this database. Grab hundreds of banana phots and train your network. It will in no time understand how a banana looks like from all angles - even angles you didn't show it before.
Just like that, we feed databases of transactions and user behavior to neural networks. We make it find patterns and let it make predictions. For example, to build a self-driving car, you need a whole bunch of sensors to give your AI digital eyes (cameras & lidar), ears (microphones), and an understanding of itself (GPS position, Accelerometer, Gyroscope) and its surroundings.
Real-world Artificial Intelligence use cases
Think about how you use Amazon. A lot of things you order allow Amazon to compare you to others and categorize you. For example, imagine you order diapers for the first time. There is a high likelihood that you and your significant other had or will have a baby. Let the baby promo start. Now it knows. A baby is on the way. It will now forever know its exact age. Always suggesting the exact right toys for every life event. And associate the persona to a future customer on that platform. Even when your little one makes their first purchase, all of their history since birth is already known.
Needless to say, having a baby is a big “money event” for Amazon so they try to predict as good as possible when that happens. It could start with you ordering a pregnancy test. Target does the same thing. To a similar degree. Rumor has it, Amazon and Target know that you are pregnant before you do.
Another example is Netflix. One of the main metrics for Netflix is the minutes you spend watching content. To drive this, it needs to recommend the exact thing you would be interested in watching. And Netflix goes all out to define this. They don't just suggest specific movies they know you might like. They alter the movie images based on what image likely resonates with you. If you like Steve Carrell and have watched “The Office” for the 10th time, it will find movies with Steve Carrell and put him on the movie poster for you. Unless there is an even stronger signal.
Think about all the transactions you make online. Search for things, buy things, sell things, read, and write things. Companies have always tried to look at consumer behavior. But never did we have the data and computing capacity available to filter useful information out of it and make predictions.
What Artificial Intelligence can do
Below is a list of other use cases companies are working on.
Anomalies in Data — Finance, Security, Risk Detection
Voice Recognition — Security, Bots, Assistant systems (Alexa)
Medical Diagnosis — Gather Research Knowledge, Symptoms Analysis based on records
Resource Planning — Manufacturing, Supply Chain Management
Business Evaluation — Accounting, Finance, Budgeting
Recommendations for Customers — E-Commerce, Media, Entertainment
Sentiment Analysis — Social Media Reputation Analysis, CRM
Image Recognition — Image Search, Face Detection, Computer Vision, Object Detection
Communication — Text translation, Real-time Speech Translation
Video Recognition — Motion Detection, Human Behavior Analysis, Safety and Security
Why are we talking about it now and not years ago?
The development and evolution of these systems have been long in the making. Algorithm design and theory started in the 80s. But only now do we have the amounts of data easily available to us. Data that is required for learning. And as stated above, the more data the better. We now have datasets big enough to train AI.
We also have the computing power necessary and affordable. These algorithms need a lot of processing power and specific processing units. GPUs make the compute efficiently and can solve these problems faster. And they are becoming more and more powerful while needing less power.
The technical breakthroughs open up massive opportunities in all industries. We will see more leaps that keep pushing Machine Learning forward. I am in for the ride and excited to see what the future holds.
Thank you for reading all the way through.
I can go deeper into details or broader and talk about other applications that are emerging. I would love to hear your feedback.
.svg)