GCP MLE: I Passed the Professional Machine Learning Engineer Exam

I had been studying to earn the Google Cloud Platform Professional Machine Learning Engineer Certification (hereafter “MLE”), and against all odds, I passed!

Apologies for an article so out of character from my usual posts, but please indulge a bit of personal bragging. That said, this should also be genuinely useful for anyone planning to take the exam.

What Is the GCP Professional ML Engineer Certification in the First Place?#

An exam where the mighty Google certifies that “this person possesses sufficient knowledge of machine learning and related GCP services.”

The exam fee is currently $200.
There are no prerequisites for taking the exam.
Currently, the exam is only available in English.
You answer 60 multiple-choice questions in 120 minutes.
The certification is valid for 2 years. Apparently, when it is about to expire, you receive a coupon to retake it for $100.

Why Did I Take It?#

I like machine learning and mess around with it a fair bit, so to test my skills
- I’ve actually posted a video like this before
I hadn’t used machine learning services on GCP much, so to learn them
Rumor has it that holding this certification makes your salary insssaaannnely high
- Rumor says the average annual salary for cert holders is $150,000 [citation needed]

Topics Covered on the Exam#

Quoting from the official certification exam guide, the following kinds of topics appear:

Section 1: Framing ML problems
Section 2: Designing ML solutions
Section 3: Designing data preparation and processing systems
Section 4: Developing ML models
Section 5: Automating and orchestrating ML pipelines
Section 6: Monitoring, optimizing, and maintaining ML solutions

That said, questions on the actual exam don’t appear neatly categorized section by section. Categorizing the questions based on my personal impression, they break down roughly like this:

Understanding the Relationship Between Machine Learning and Business#

Even before machine learning itself, you are expected to have common sense about how to solve a given problem in the first place.

Machine learning is not a silver bullet that solves every problem.
If a problem can be solved without using machine learning, do it that way.
Even if machine learning is a good fit, if you haven’t gathered enough data yet, there’s no need to introduce machine learning right from the start.

As long as you firmly take this Google philosophy to heart, you’ll be fine.

General Knowledge of Machine Learning#

Before even considering GCP services, general knowledge of various machine learning algorithms, analytical techniques, and statistics is required.

There are various types of neural networks, but what is each one good at?
- Images: CNN
- Time-series data: RNN
- Tabular data: Standard neural networks
What is overfitting? What are the signs? How should you deal with it?
- An overly complex model failing to generalize.
- Training loss is tiny, but test loss is ridiculously large, etc.
- Increase the data, use L2 regularization, simplify the model, add dropout layers, and so on.
What are Precision and Recall? Can you use ROC AUC on imbalanced data?
- Precision is “of all the samples predicted as positive, how many were actually positive.”
- Recall is “of all the samples that were actually positive, how many were predicted as positive in advance.”
- For imbalanced data, evaluate your model using PR ROC.

And so on and so forth.

Undergraduate-level knowledge is basically enough, but be careful because questions are often presented in the context of business situations.

A Sample Question#

As the machine learning lead, you are trying to optimize the maintenance frequency of a factory.

You decide to build a model that predicts which machines are about to break down, and perform an inspection on any machine predicted to fail soon.

Inspections don’t cost very much, but slacking on inspections and letting a machine break down is insanely expensive.

What metric should you use to select the model?

A. The model with the highest Accuracy.

B. The model with the highest Recall.

C. The model with Precision of 50% or higher and the highest Recall.

D. The model with Recall of 50% or higher and the highest Precision.

Answer#

Because “letting a machine break down is insanely expensive,” you want to minimize missed failures as much as possible. That means you want high Recall. Therefore, C is the correct answer.

A and B are eliminated because they could result in sloppy models that just predict all positive or all negative.

If you choose D, you would be trying to minimize cases where “we called for an inspection, but it wasn’t actually about to break.” That doesn’t match the objective, so it’s eliminated.

Vertex AI
AutoML
BigQuery ML
Kubeflow
Dataflow
Dataproc
Dataprep
Data Fusion
TensorFlow
TensorFlow Extended (TFX)
TensorFlow Data Validation

And so on. Since this is a GCP exam, this is naturally the main event. Even if not directly related to building models:

Where to store data
- BigQuery
- Cloud Storage
- Bigtable
How to ingest data
- Pub/Sub
- Cloud Functions
CI/CD
- Cloud Build
- Cloud Source Repositories
- GitHub

You need to know the features, pros and cons, and even the syntax and commands for all of these various services required for MLOps.

A massive number of questions take the form: “You want to achieve X. There is constraint Y. You want to minimize Z. Which combination of services should you use?” You end up using a process of elimination following a flow like this:

Can That Combination Actually Achieve X in the First Place?#

Annoyingly, however, quite often all four options are things that “could technically be done if you really wanted to.”

Does It Comply with Constraint Y?#

Pay close attention to phrasing like without writing code or no team members have machine learning experience.

If it says without writing code, you must not choose an option that uses TensorFlow.

If it says no team members have machine learning experience, basically picking AutoML is fine.

Does It Follow Google’s Recommended Best Practices?#

If it takes some nonsensical detour, it’s out.

For example, for processing data in BigQuery and storing it back in BigQuery, using anything other than BigQuery is basically a no-go.

However, if the question prompt happens to say something like without writing code, then an option configuring Dataflow using Dataprep instead becomes justified.

Of the Remaining Options, Which One Minimizes Z?#

Whether it’s cost, training time, prediction latency, manual human effort, or required coding knowledge. When you’re left with services that can do similar things, you’ll be stuck here unless you also know the pros and cons of each one.

Study Routine and Tips for Anyone Aiming for the Same Certification#

After learning on Coursera, I solved practice problems over and over.

Coursera#

First, spending about two months, I completed the entire Coursera program: Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate.

The course is mostly videos, but hands-on “try it yourself” segments (Labs) are sprinkled in throughout. They lend you a temporary GCP account with a time limit, which is wonderful because you can play around with services that normally cost a fair bit of money, completely for free. And getting hands-on experience really deepens your understanding.

Udemy Practice Exams#

Over the next month, I hammered through practice exam questions on Udemy from Google Cloud Professional Machine Learning Engineer 2022.

It includes 170 questions: 50 questions per test across 4 tests (only the last test has 20 questions).

It’s really nice that each question comes with fairly detailed explanations. The question format and exam trends are almost identical to the real thing, so it makes for great practice.

Note, however, that newer services and open-source tools are not covered.

The course description recommends repeating until you consistently score 90% or higher on each test, but remember it’s just a practice exam—the exact same questions won’t show up on the real test.

Rather than obsessing over your practice exam score, you should focus on reading the explanations and reference materials thoroughly to deepen your knowledge of each topic.

Tips for Taking the Exam#

Other useful information.

When Registering Your WebAssessor Account, Register Your Name in Japanese#

If you register as “Yuko Yoshida” (in the Latin alphabet), you’ll need to present government-issued ID that has your name written in English—in short, a passport.

If you register as “吉田優子” (in Japanese characters), standard ID like a Japanese driver’s license will be accepted.

For details, see Acceptable Identification.

If by any chance you registered with an English name without having a passport, you’ll need to contact support.

Why do I know this in such detail? Because I did this exact screw-up myself.

You Can Take the Exam at Home or at a Test Center#

For home testing, you apparently take the exam while connected to a proctor via Zoom. I took mine at a test center.

Test Center Reservations Can Be Canceled or Rescheduled Free of Charge Up to 72 Hours in Advance#

Weekend slots for the upcoming two weeks are usually booked up, so it’s a good idea to just go ahead and reserve a slot. If you find yourself unprepared right before the exam, you can simply reschedule then.

By the way, for home testing, I believe rescheduling becomes locked 1 week in advance. Definitely verify this.

Read the Exam Guide Email Carefully#

You will need two forms of identification. Prepare them in advance.
Arrive at the test center 15 minutes before the exam start time.

Mindset During the Exam#

Don’t get discouraged even if difficult questions come up. Easy ones are mixed in too.
- In my case, I couldn’t figure out about the first 5 questions in a row, and my spirit nearly broke.
For questions you don’t know, flag them with Review Later and move on quickly.
- Maybe because the tension eased up or my brain switched into English mode, but on the second pass, the question prompts flowed into my head much more smoothly.

In my case, I checked Review Later on every single question where I lacked absolute confidence.

At the end of round 1: 40 minutes remaining, with 30 questions marked for review.
For round 2, I revisited only the flagged questions. Although I couldn’t make it all the way to the end, game was called right after I managed to whittle the flags down by 10.

That’s how it went.

Mindset When Finishing the Exam#

When the exam ends, there’s a survey of about 12 questions taking roughly 10 minutes, but this is optional. You can answer it or skip it.

After that, without any fanfare or fanfare-like emphasis, a simple message like Result: Pass pops up nonchalantly.

The word Pass was in red text, so my heart stopped, but after rereading it about three times, it finally sank in: “Hell yeah!”

Mindset After the Exam#

The exam result is reflected in WebAssessor right away (strictly speaking, it’s a provisional result until Google’s official approval comes through).

Check your results again and bask in the glory.

In Conclusion#

Apparently, you get certification swag. Looking forward to it.

Addendum#

My certificate arrived! It took less than a week from exam day.

The certification swag currently seems to be a choice between a vest and a charity donation. I think I’ll write another post about that whole story sometime.