What I learned studying for the GCP Machine Learning Exam
04 Nov, 2020 Read Time: 12 Minutes
In August 2020, Google launched a BETA version of a new addition to their GCP certification library: Professional Machine Learning Engineer. This Certification spans a broad range of machine learning topics including design, implementation, deep-learning frameworks and documentation.
I always like to challenge myself with Google Cloud, so this BETA certification was too good an opportunity to pass up. Plus I want to be considered an expert in Machine Learning with Google Cloud Platform, so it was time to put that into practice.
Being one of the first to pass the exam in October, along with my Colleague Di Wu, I wanted to share my thoughts on this certification and how you should prepare. So here are a few takeaways from my experience.
Firstly, for the uninitiated, what does a Machine Learning Engineer do?
We build and optimise machine learning models, primarily with the purpose of taking a product from first trial size to end-user. The initial model is taken from a Data Scientist, who is more a jack of all trades and will build the model and ensure it actually works.
Although similar, there’s a key difference in these two roles. Data Scientist’s don’t look at productisation or building a complete solution; their role stops with the model. The Machine Learning Engineer takes extra steps to improve the model, refining and updating it and taking into account all product and performance recommendations so it’s up to scratch.
What do you need to do to pass?
If you’re interested in gaining expertise within Google Cloud technology, this certification is for you - but be warned, Machine Learning is by far the toughest Google exam I’ve taken so far.
To pass you have to have good software principles, a strong knowledge of the machine learning capabilities of the GCP stack and Machine Learning & GCP theory - which other Google certifications haven’t really included so far. The Machine Learning certification is valid for 2 years and definitely worth having to stand out in the field.
What do I need to study in depth?
There are a broad range of topics within the Machine Learning exam - but I’d say the key things to brush up on are:
- Cloud technologies and software
- CI/CD & Pipelines, BigQuery, Datalab and other tools in the GCP stack
- Software design principles
- Continuous delivery vs. continuous development
- How to test and optimise model performance (the models vary so it’s good to understand classification, linear regression, etc.)
- Moderating performance and how to understand performance metrics (eg. precision vs. recall vs. f1 score)
- TensorFlow and KubeFlow development framework
- Solution architecture (i.e. when to use a custom solution vs. a package solution)
- Understanding the ethical uses of AI (user privacy is key)
Finally, you also want to have a deep understanding of business use cases; applying the business side to machine learning and fair valuation, too.
You can view the full list of topics I encountered at the end of this article.
How much experience do you need to pass?
I’d say you need at least 2 years general GCP experience to take the exam. As for hands-on machine learning experience? It’s tough to say; if you know the topics, then you’ll be good to go. Some have passed with just 6 months experience. Having said that - at least a year of professional experience developing solutions in the market, not just looking at theory, would be helpful. A data engineering certification would also be valuable to have in your arsenal, but not essential.
Preparing for the Google Machine Learning exam
You need to know how to monitor tests - and how you evaluate model performance is important. The math behind machine learning - what the different model parameters do, algorithm optimising, etc - are all helpful for this exam. I’d also recommend studying TensorFlow data structures and Tensorflow’s DataSet API. Finally, look at how to optimise solutions from a technical perspective as well as a business perspective.
Before taking Machine Learning, consider the other certifications
While you can complete this exam in isolation I’d recommend completing other Google Cloud Certifications, starting with the Associate Cloud Engineer. Similar topics are covered, so it’d help with general awareness and knowledge.
I’d also recommend the Professional Data Engineer exam, although it’s not absolutely necessary. If you’ve also done the Cloud Engineer certification you’d need to study less as you already know about data storage, etc. If you can, get training on the Machine Learning certification and the GCP platform.
What are the toughest parts of the exam?
Definitely the math side; the test is online-proctored due to COVID-19, so you can’t use a pen or paper. That’s difficult as it’s all off the top of your head. Although it’s worth noting that, as a Beta exam, the math elements may not be included in the final version.
You also need to be up to date on documentation - the function or concept may not have changed but the documentation may have been updated and the name of concepts also updated.
Lastly, the software development side can be difficult - in particular, good design principles and developing continuous pipelines. So brush up on those bits!
If I were sitting this exam again, I’d have studied the documentation more
They’ll expect you to be up to date with the latest Google technology and terminology. Ditto the TensorFlow data structures and concepts. I’d also be more cautious of software design principles.
I only had two weeks to study so more time building solutions would’ve been helpful - I’ll say it again, if you can get some GCP Training that’ll help for sure.
My final tip? Get online-proctor ready
It’s a remote exam - for now - so have the room clear, no notes, books or noise in the background so your results aren’t invalidated. Make sure you have a strong internet connection so there are no distractions or disruptions. If you have time, try to build some of the concepts out to make life easier for yourself. Finally, use deductive reasoning to eliminate any wrong answers and read the questions very carefully to help you with any multiple choice answers.
Good luck!
Appendix 1: Exam basics
https://cloud.google.com/certification/machine-learning-engineer
Length: Two hours
Registration fee: $200 (plus tax where applicable)
Languages: English
Exam format: Multiple choice and multiple select
Exam Delivery Method:
- Take the online-proctored exam from a remote location, review the online testing requirements here.
- Take the onsite-proctored exam at a testing center, locate a test center near you.
Prerequisites: None
Recommended experience: 3+ years of industry experience including 1+ years designing and managing solutions using GCP.
Appendix 2: Topics I encountered in the machine learning exam
The questions in this exam covered a wide range of topics including some very simple programming questions:
- How to improve the performance of a model
- How to improve the performance of hardware
- Some data engineering questions
- TFX & Kubeflow Questions
- Cloud Build
- BQML, Tensorflow & Sci kitlearn
Tensorflow topics:
- Tensorflow Records
- Tensors
- TFX (Tensorflow extended)
- TPUs
- When to use them
- How to improve them
- What are estimators
- How to deal with common errors
- Data API
Kubeflow topics:
- Why use kubeflow
- Kubeflow components
- Benefits of using Kubeflow
- How to automate Kubeflow
- What is Kubeflow hybrid
Machine Learning:
- When to use supervised learning vs unsupervised
- Decision Trees, TF models, Transfer learning
- When would you use TF
- Evaluating models
- Recall
- Precision
- F1 score
- How to solve overfitting and underfitting
AI platform:
- How it works
- How to store models
- Uploading data to it
- Notebooks
- Monitoring
- How you can use it with R
- How it can be used to work locally and on the cloud
- Improve performance
- Training on cluster
CI/CD & Pipelines:
- What is continuous ingestions and how you can use GCP with it
- What is continuous delivery and how you can use GCP with it
- How to use Cloud Monitoring to ensure pipelines work
- How to use Tensorflow to ensure pipelines work
- How to use Kube Flow to ensure pipelines work
- How to use other GCP monitoring tools Stackdriver, Logging etc
- When you want to use a pipeline a certain way
- How to deal with low latency
- How to deal with slow performing pipelines
- Building real-time and batch systems
Training & Testing:
- How to avoid training skew
- How to use preprocessing functions in TF to help with training and export that to testing and live predictions
- How TFX can help
- How to do this across different tools like AI platform, BQML etc
BQML:
- What can you do with it
- What algorithms exist
- When can you use it
Hardware:
- What are TPUs and how they work
- What are CPUs and how they work
- What are GPUs and how they work
- When to use each one
- How to use TPU 3.0
- How to use GPU Nvidia
- How to use CPUs
- How to deal with lack of memory issues
Business and ML:
- How can you use ML to help solve this problem
- Typical questions: You run a delivery app company and how can you empower your drivers by using ML. You have data from this source and this source and typically drivers complain about X
- You want to increase revenue by suggesting related products what approach do you suggest
- Evaluating the model performance against business metrics:
- How do you devise a testing plan to see model performance
Pricing, Permissions and Privacy/governance:
- Responsible AI practices
- Costs/benefits of using different GCP products
- How to provide permissions to the model
Evaluation:
- How to build a process using different GCP tools to ensure your model performance is always good
- How to use AI explanations to help understand model performance
- How to use TFX to evaluate model performance
- How to use Kubeflow
Hybrid Cloud Models:
- How to work with hybrid cloud system
- What to do on private cloud vs public
- How Kubeflow can help











