Azure AI Fundamentals: How to Pass the AI-900 Exam

I was speaking to my dear 78-year-old mother today, and she was telling me that Microsoft just bought an Artificial Intelligence (AI) company. Now I would never in a million years suggest septuagenarians are out-of-touch with modern technology, but it still surprised me she would bring this up. 

Here’s the interesting thing that ties a 78-year-old with AI—the company that Microsoft bought (called Nuance) makes AI software in the medical dictation field. My mother worked in medical dictation when she first moved to Canada at 19.

So after my mother told me that Microsoft had bought this company, I had to tell her that this company makes smart computers that do the job that she used to do. Very few medical transcriptionists are employed today compared to the 1960s.

Here is the Google Trends for the term “medical transcriptionist” over time:

The above graph reflects a type of job that AI has largely replaced.

AI is everywhere in our daily life, in places we can see, and in places that we cannot. When we talk to our phones and our smart speakers to tell them to perform some task, a computer in a far-away data center interprets the sound and determines our intention. That computer can understand our speech and our subtle intention without someone having programmed those specific words to mean that specific thing.

“Alexa, it would be really unbelievably amazing if you could turn down the lights right now. Thanks, Hon!”

And almost at the same time, in a different computer system, an AI computer algorithm has decided that our creditworthiness has improved and is preparing to send us a letter offering us a higher credit limit. This is automated machine learning (ML) without any human decision or intervention.

Some might say that AI and ML are our future, but I say that it’s our present. AI and ML are everywhere. They are in our cars, in our traffic lights, in our homes, in our pockets, and in our workplaces.

If a computer wrote this blog post, you might not even know it. There’s an interesting ML program called GPT-3 that can write articles in a way very similar to how humans write. 

If AI interests you at all, whether you have a technical background or not, perhaps you should grab a course on it and learn the basics. One way to do that is to take a basic AI certification test like the one Microsoft offers.

The AI-900 exam from Microsoft tests a wide range of basic foundational knowledge when it comes to general AI and ML principles. It also ensures that the candidate has basic knowledge of Microsoft Azure’s service capabilities when it comes to AI.

Cost: US$99

Questions: 40-60

Time Limit: 60 minutes (90 if you include the surveys before and after)

Question Types: Multiple Choice

Experience Level: Beginner

Certificate Expires: Does not expire, good for lifetime

Now maybe you’ve heard of AI before, but what actually is it? Are we talking about a futuristic computer that goes back in time to track down John Connor? No, not that kind of AI. That’s called Generalized AI and has not been invented yet, to our knowledge.

When we say “AI,” we’re talking about a very specialized type of AI. That specialized AI has  very limited and specific training and capability. Data scientists train a specialized AI through a technique called ML. 

Loosely, you could define AI as any technology that aims to replicate the human brain. A human brain takes input from the five senses: 

• Visual input through the eyes
• Audio input through the ears
• Olfactory input through the nose
• Tactile input through touch
• Gustatory input through taste

The human brain uses those inputs to help it make decisions and take action. It can understand conversations happening around it, reply to those conversations, make predictions about what is about to happen based on what it sees, or just recognize certain objects. 

Now it’s not common for a cloud computer to have a nose, touch, or taste, but AI takes input from its senses:

• Video input
• Audio input
• Textual input

Based on those inputs, an AI brain also makes decisions and can be programmed to take actions.

Common AI workloads and important considerations

For the AI-900 Azure AI Fundamentals exam, it’s important to understand the common AI workloads. Using these inputs, an ML algorithm can:

• Make forecasts or predictions (such as weather forecasts)
• Classify objects (is this a penguin or a cat)
• Arrange data into clusters (these items are somehow related)
• Detect anomalies in a stream of computer data
• Translate text between languages
• Synthesize speech
• Have a conversation with a human

Having a computer that can make important decisions without human intervention or override is a dangerous concept. Movies have been made about the consequences of that. (“I’ll be back!”) So Microsoft has developed guiding principles for responsible AI. These are tested on the exam as well.

When people train an AI model to perform a task, they do it with the best of intentions. Many HR departments across the world would like a computer to filter through all of the job applicants they get for any job and automatically filter out candidates that do not meet certain criteria to be successful.

With AI solutions, you do not program the exact criteria. You ask the computer to examine the past successful (and unsuccessful) applicants, and figure out what they all have in common. Then you can pre-determine which candidates are likely to be hired and which are not.

Unfortunately, what if your HR resume filtering program determines that only men are likely to get hired? So it starts filtering out people with non-male-sounding names before sending the resumes to HR. You’ve just automated an anti-female bias.

This is not far-fetched. This actually happens. 

That’s why Microsoft’s fairness principles talk about the importance of fairness, reliability, safety, privacy, security, inclusiveness, transparency, and accountability.

The AI-900 exam does ask about these subjects as well.

Principles of machine learning on Azure

The AI-900 exam is a fundamental-level exam. You are not required to actually design and implement solutions using Azure ML and AI services.

However, it’s important to have a good grasp of the options that Azure provides. You should have a handle on the terms and how ML works. You should be able to describe to someone if an ML model is a good one or not.

For example, here are the most common types of ML algorithms:

• Regression algorithms: used for prediction
• Classification algorithms: used to identify an object
• Clustering algorithms: used to group related items together
• Deep learning: using artificial neural networks

Successful ML starts with lots of good data. The dataset you provide to the machine needs to be a fair representation of what the algorithm will encounter in the real world. No group should be too under- or over-represented. You might think that 10,000 rows of data are a lot, but 100,000 rows are better, and a million rows are better than that.

The first stage of developing an ML model is data ingestion and preparation. You need to get the data from wherever it is into Azure. That data might not be formatted well, and maybe there is “bad data” in it. You need to clean it up.

The second stage of developing a model is feature engineering and feature selection. You might have 200 columns of data on each row, but it’s unlikely that you need all of it. Perhaps only a handful of columns are relevant to the decision you are trying to teach the computer to make. If you’re trying to predict future product sales, the customer’s location might be relevant, but surely the name of the customer who bought it is not.

“Data shows 12% of our sales are to people named John, so we are starting a Facebook ad campaign only for people named John.”

The third stage of developing a model is the actual training and evaluation. You direct the model to the data, using the selected algorithm, and let it process all of the data. The key is to then evaluate the model by having it predict the outcomes of a set of data it has never seen before. You can then see how accurate it is when using the testing data.

Finally, the last step of developing an ML model is to deploy the winning model to production and begin to use it in your real application.

Computer Vision workloads in Azure

Another topic on the AI-900 exam is Azure Cognitive Services. Cognitive Services is a pre-built set of ML models that you can use without having to train them. However, you can sometimes customize them with your own data.

One set of services is called Computer Vision. This application programming interface (API) can be used to analyze the contents of video and images, and extract valuable information from them. For instance, Computer Vision API can:

• Extract handwritten and printed text from images
• Recognize celebrity faces
• Recognize worldwide landmarks
• Recognize common objects such as a bicycle or a phone
• Recognize where human faces are in a photo
• Estimate the age and genders of the people in a photo
• Find the same person in multiple photos
• Describe a photo
• Extract tags from a photo
• Extract the field contents from a form or invoice
• And many more tasks

It’s really quite impressive when you see how good computers are at recognizing the contents of images these days.

Natural Language Processing (NLP) workloads in Azure

When a computer can understand the meaning and intention of human words, that is called “natural language processing” or NLP.

This technology has made searching the Internet so much easier. You can enter some words into your favorite search engine like Google, and the site will do a much better job at guessing what you were trying to find than just simple keyword matching. 

Computers have also gotten much better at translating text between languages. Even 10 years ago, translating text was more of a literal translation. In the early days of the Internet, it was a popular joke to translate text from English to another language, and then back to English again, and laugh at the result. Nowadays, you are more likely to discover that doing this results in a fairly accurate translation. It’s not funny anymore.

Beyond understanding written words, spoken words, and translation, computers are also getting better at understanding meaning. Computers can now read lengthy written text, (such as this blog post), and summarize it into a few words fairly accurately.

Conversational workloads in Azure

The last type of AI workload covered by the AI-900 exam is conversational AI. This sometimes takes the form of chatbots. These are the “friendly” pop-ups that you sometimes see when looking at company support pages. Ask your question in your natural language, and the chatbot will try to find the answer for you without having to connect you to a support agent.

“What time are you open until tomorrow night?”

The computer can understand what you are asking and will return the store hours of operation for tomorrow.

You can build your own chatbot using the Azure Bot service. More than just answering basic questions, you can develop a workflow in Azure. In fact, the bot can take an order over chat. You could order a pizza using a chatbot, and the AI chatbot technology in the backend makes sure it has understood your order correctly.

Summing it all up

Those are the top-level subjects covered by the AI-900 Microsoft Azure AI Fundamentals exam. You do not need to be a developer or an expert in these services at all—just a basic understanding of what they do.

If you’re interested at all in taking the AI-900 test, I think you can do it with only a couple of weeks of study. I have an AI-900 course on Udemy that covers all of the topics of this exam. The course is covered by Udemy’s unconditional 30-day money-back guarantee, but I am confident that you will not need to use that.

Hundreds of students have taken this course and used it to pass the exam. I get “thank you” messages every day from them. You can read them in the reviews. I am confident you can pass this exam with a bit of effort. Challenge yourself!
Other Azure exams that complement the AI Fundamentals Exam include the AZ-900 Azure Fundamentals Exam and the DP-900 Azure Data Fundamentals Exam.