AutoML and Low-Code ML with Vertex AI – they solve the problem that building machine learning models traditionally required deep programming skills and weeks of trial and error. For the PMLE exam, this is crucial because Google wants you to know how to rapidly create production-ready models even if you are not a coding wizard, and how to choose between AutoML and custom training depending on the business need.
Jump to a section
A simple way to picture AutoML and Low-Code ML with Vertex AI
You want to cook a great meal for a friend who loves Italian food, but you have never chopped an onion or boiled pasta in your life. What do you do?
You could spend months learning to be a chef – studying knife skills, sauce temperatures, and plating techniques. Or you could hire a personal chef. You tell the chef what you want: 'A classic spaghetti carbonara, but my friend is allergic to dairy and can't eat pork.' The chef handles everything – selecting ingredients, cooking, tasting, and adjusting – and delivers a perfect meal. You never touched a pan.
AutoML and low-code tools in Vertex AI work exactly like that personal chef. In traditional machine learning, you are the chef: you write all the code, clean the data, choose the algorithm, train the model, and tweak it until it works. That takes months and deep expertise. With AutoML, you are the customer. You tell the tool what problem you want solved – for example, 'identify whether a customer review is positive or negative' – and provide the data, like past reviews. Vertex AI AutoML automatically tries dozens of different model architectures, fine-tunes them, and presents you with the best performing one. Low-code tools like Vertex AI Pipelines and the Workbench let you drag and drop components instead of writing thousands of lines of code. They give you the power of a professional chef without requiring you to become one.
Let's start with a clear definition. Machine learning (ML) is a way for computers to learn patterns from data without being explicitly programmed for every rule. For example, you do not write code that says 'if the email contains the word 'free' and 'money', mark it as spam'. Instead, you show the computer thousands of emails that are already labelled spam or not spam, and it figures out the patterns itself.
Building that model traditionally involves several steps. First, you have to collect and clean the data – that means removing duplicates, handling missing values, and making sure the data is in a consistent format. Then you have to choose an algorithm – like a decision tree, a neural network, or a support vector machine. Each algorithm has dozens of settings called hyperparameters that need to be tuned, like the learning rate or the number of layers in a neural network. Then you train the model by feeding it data, evaluate it to see how accurate it is, and then iterate – change settings, retrain, test again. This process, called model development, can take weeks or months and requires a data scientist who knows Python, TensorFlow, or scikit-learn.
AutoML (Automated Machine Learning) changes this completely. AutoML is a suite of tools that automates the most difficult parts of the ML workflow. You still need to provide high-quality labelled data – that is the raw ingredient – but Vertex AI AutoML handles the algorithm selection, hyperparameter tuning, and model evaluation automatically. It uses a process called neural architecture search, where it tries many different model structures in parallel, picks the best ones, and refines them. At the end, it gives you a model you can deploy directly, without writing a single line of code.
Low-code ML takes this a step further by providing visual interfaces for the whole pipeline. Vertex AI includes tools like the Vertex AI Pipelines UI, where you can drag and drop steps like 'ingest data', 'preprocess', 'train', 'evaluate', and 'deploy' onto a canvas and connect them. The tool automatically generates the underlying code. Another low-code tool is the Workbench, a managed notebook environment that comes with pre-built code snippets and templates so you can build models with minimal scripting.
Why does this exist? Three main reasons:
It democratises ML: Not every company can hire a team of PhD data scientists. AutoML puts ML in the hands of business analysts and software developers.
It speeds up development: Instead of waiting weeks for a model, you can have a prototype in hours.
It reduces human error: Tuning hyperparameters manually is error-prone. AutoML systematically searches for the best combination.
What does AutoML replace? It replaces the traditional custom model building workflow. In a custom workflow, a data scientist writes code using a framework like TensorFlow or PyTorch, manually splits data into training and testing sets, chooses an algorithm, manually tunes hyperparameters, and iterates. AutoML replaces all of that manual coding and trial-and-error with automation. However, it does not replace the need for good data. Garbage in, garbage out still applies. If your training data is biased or incomplete, your AutoML model will be biased or inaccurate.
Here is a concrete example. Imagine a bank wants to build a model that predicts whether a loan applicant will default. With traditional ML, a data scientist would spend weeks writing Python code, trying logistic regression, then random forests, then gradient boosting, testing different feature combinations, and tuning parameters. With AutoML, the bank uploads a CSV file with historical loan data – columns like income, credit score, loan amount, and whether the applicant defaulted. Vertex AI AutoML automatically analyses the data, tries dozens of model types, picks the best one, and outputs a ready-to-use model. The bank's analyst, who may not know Python, can do this through the console in a few hours.
Low-code tools extend this to the entire ML pipeline. For example, Vertex AI Pipelines allows you to build a repeatable workflow that runs automatically whenever new data arrives. You can set up a pipeline that ingests data, runs an AutoML training job, evaluates the model, and deploys it to an endpoint – all with a visual drag-and-drop interface. This is called MLOps (Machine Learning Operations), and it is a critical concept for the PMLE exam.
To summarise the key technical terms:
AutoML: A service that automatically trains, tunes, and evaluates the best model for your data.
Low-code: A visual development environment that reduces the amount of manual coding needed.
Vertex AI: Google Cloud's unified platform for building, deploying, and managing ML models.
Hyperparameters: Settings that control how a model learns, like the learning rate or number of layers.
Neural Architecture Search (NAS): The automated process of trying different model designs.
MLOps: Practices for deploying and maintaining ML models in production.
In the PMLE exam, you will be tested on when to use AutoML versus custom training, how to set up AutoML training jobs, and what low-code tools are available in Vertex AI. You will also need to understand the limitations: AutoML works best for standard tasks like image classification, text classification, and tabular data. For cutting-edge research or highly specialised architectures, custom training with code is still necessary.
Prepare and upload data
Collect clean, labelled data in a supported format (e.g., CSV for tabular, images in Cloud Storage for image). Upload the data to Google Cloud Storage. This step is crucial because data quality directly determines model performance – AutoML cannot fix bad data.
Create a Dataset in Vertex AI
In the Vertex AI console, create a new 'Dataset' of the appropriate type (tabular, image, text, or video). Import your data from Cloud Storage or BigQuery. The system will automatically split it into training, validation, and test sets. This step organises your data so AutoML can use it.
Train an AutoML model
From the dataset, click 'Train New Model' and select 'AutoML' as the training method. Configure the objective (e.g., classification, regression) and optionally set a training budget (max node hours) to control costs. Vertex AI automatically tries multiple architectures and returns the best model.
Evaluate the model
Review the model evaluation metrics provided by the job – precision, recall, F1 score, confusion matrix, and feature importance. These tell you how accurate the model is and where it makes mistakes. Accept or reject the model based on business requirements.
Deploy the model to an endpoint
Create an endpoint in Vertex AI and deploy the model to it. You can choose a machine type (e.g., GPUs) and set scaling policies. The endpoint provides a REST API URL that applications can call to get predictions in real time.
Use predictions and monitor
Send prediction requests to the endpoint URL from your application. Monitor the model's performance over time using Vertex AI Model Monitoring, and set up scheduled retraining pipelines with Vertex AI Pipelines to keep the model current as data evolves.
Meet Priya, a data analyst at a mid-sized fashion retailer called TrendVibe. The company has an online store with thousands of products, and customers frequently upload photos of themselves wearing outfits. Priya's boss asks her to build a system that automatically tags uploaded photos with the type of clothing shown – 't-shirt', 'jeans', 'dress', 'jacket' – so the marketing team can analyse trends.
Priya has never written a line of Python. She knows Excel well and understands the business need, but building a machine learning model from scratch seems impossible. Here is what she actually does using Vertex AI AutoML:
Prepare the data: Priya works with the marketing team to collect 10,000 customer photos that are already labelled with clothing types. She uploads the images and a CSV file mapping each image to its label (e.g., 'img_001.jpg', 'dress') into Google Cloud Storage, which is a cloud-based file storage service.
Create a dataset in Vertex AI: In the Google Cloud console, Priya navigates to the Vertex AI section and clicks 'Create Dataset'. She selects the type 'Image' and the objective 'Single-label classification' – meaning each image gets exactly one tag. She then imports the images and labels from Cloud Storage. The interface shows her a preview of the data, and she checks that labels are balanced – roughly 2,000 images per category.
Train an AutoML model: Priya clicks 'Train New Model'. She selects 'AutoML' as the training method, gives the model a name, and clicks 'Start Training'. Vertex AI automatically chooses a neural network architecture, splits the 10,000 images into training, validation, and test sets, and begins training. The console shows a progress bar and estimated time – about three hours. She walks away to attend a meeting.
Evaluate the model: When training finishes, Priya receives an email notification. She opens the model evaluation page, which shows metrics like precision, recall, and an overall accuracy of 92%. A confusion matrix shows which categories the model most often confuses – for example, 'jeans' and 'trousers' are mixed up 5% of the time. She decides this accuracy is good enough for the marketing team's needs.
Deploy the model to an endpoint: Priya clicks 'Deploy to Endpoint'. She chooses a machine type with one GPU – the tool offers recommendations based on expected traffic. Vertex AI creates a REST API endpoint, which is a web address that accepts image uploads and returns predictions.
Integrate with the app: TrendVibe's app developer takes the endpoint URL and writes a simple integration – a few lines of code – so that whenever a customer uploads a photo, it is sent to the endpoint and the tag is stored in the database. Priya did not need to know any ML algorithms to make this happen.
What did Priya do that a data scientist would have done differently? If Priya were a data scientist, she would have spent weeks writing Python code to preprocess images, build a convolutional neural network, tune hyperparameters, and handle edge cases. Priya's AutoML approach took one day of data preparation and three hours of automated training. The trade-off is that Priya has less control over the inner workings of the model, and if the marketing team later needs a custom architecture that AutoML cannot generate, she would need a data scientist. But for 90% of common business problems, AutoML is sufficient.
This scenario illustrates exactly what the PMLE exam expects you to understand: how to use Vertex AI AutoML for tabular, image, text, and video data; when to choose AutoML over custom training; and how to interpret the evaluation metrics that AutoML provides.
The PMLE exam tests your understanding of AutoML and low-code tools in two major question categories: scenario-based multiple choice and direct knowledge recall. The exam expects you to make decisions about when to use AutoML versus custom training, how to configure AutoML jobs, and what the limitations are.
Here are the exact concepts they test and the traps they set:
AutoML vs Custom Training: This is the most tested distinction. You will get a scenario – e.g., 'A company has 50,000 images of products and wants to build a classifier. They have a team of business analysts with no ML experience. Which approach?' The correct answer is always AutoML because it requires no coding. The trap: They describe a scenario where the model needs to use a newer architecture not yet supported by AutoML, in which case the answer is custom training with code.
Data types supported by AutoML: You must memorise that Vertex AI AutoML supports tabular data, image data, text data, and video data. For tabular, it supports regression, classification, and forecasting. For image, classification, object detection, and segmentation. For text, classification, sentiment analysis, and entity extraction. A common exam trap is asking about whether AutoML supports generative AI or large language models – it does not natively; that is handled by Vertex AI Generative AI Studio.
Managed dataset creation: The exam tests that you know a Dataset in Vertex AI must be created before training. The dataset can be imported from Cloud Storage with a CSV or from BigQuery. Traps: They ask about importing data directly to a model – that is wrong, you need a dataset first.
Training options within AutoML: You can choose between 'AutoML', 'Custom', or 'AutoML Edge' for on-device models. The trap is offering 'AutoML Lite' – that is not a real thing.
Explainable AI with AutoML: AutoML models support feature importance explanations, including tabular data feature attributions and image model explanations. The exam tests that you can retrieve explanations via the API or console. Traps: They may suggest that AutoML models are black boxes with no explainability – that is false.
Low-code tools list: You need to know the specific low-code tools in Vertex AI – the Workbench (managed notebooks), Vertex AI Pipelines (drag-and-drop orchestration), Vertex AI Model Registry (model versioning), and Vertex AI Experiments (tracking experiments). A trap: They list tools from other Google Cloud products like Cloud Composer or Dataflow and ask which is a low-code ML tool. Cloud Composer is for workflow orchestration, but it is not specifically an ML low-code tool.
Limitations of AutoML: AutoML works best with datasets between 1,000 and 1,000,000 samples. For tiny datasets, it might overfit. For highly specialised models like transformer-based NLP models, custom code is required. The trap: A question implying AutoML can handle any dataset size or any model type.
Pricing and compute: AutoML charges for compute hours and training time. You can set a budget (max node hours) to control costs. The exam may test that you know how to limit costs by setting a budget.
MLOps with AutoML: You can use Vertex AI Pipelines to automate AutoML training, evaluation, and deployment. The exam tests that you know how to create a pipeline that triggers AutoML training on a schedule or when new data arrives.
Example exam-style question: 'Your company needs to build a model that predicts customer churn using a CSV file with 20,000 rows. The data scientist is on leave, and you need a model by tomorrow. What should you do?' Correct answer: Use Vertex AI AutoML for tabular data classification. Wrong answer: Write a custom TensorFlow model or wait for the data scientist.
The key definitions you must memorise:
AutoML dataset: A managed dataset that contains your training data, labels, and splits.
Training pipeline: The automated workflow that includes data preprocessing, model training, and evaluation.
Model evaluation: Metrics like precision, recall, accuracy, F1 score, and confusion matrix that AutoML computes automatically.
Endpoint: The deployed model that accepts prediction requests.
Explainable AI: Tools that show which features most influenced a prediction.
The exam loves to test edge cases: what happens if you upload a dataset with incorrect labels? The model will have poor accuracy. What if you need a model that runs on a mobile device? Use AutoML Edge. What if you need real-time predictions? Use an endpoint with a small machine type. What if you need batch predictions? Use a batch prediction job.
Finally, be aware that the PMLE exam is evolving and now includes questions about generative AI and foundation models, but those are covered in a separate exam section. For section 6.1, focus purely on AutoML and low-code tools – do not confuse them with Vertex AI's Generative AI Studio or Model Garden.
AutoML in Vertex AI automates algorithm selection, hyperparameter tuning, and model evaluation, requiring no manual coding for standard ML problems.
Low-code tools in Vertex AI, including Pipelines and Workbench, provide visual interfaces that generate underlying code, reducing development time.
You must prepare clean, labelled data before using AutoML – data quality is the user's responsibility, even though training is automated.
Vertex AI AutoML supports four data types: tabular, image, text, and video, each with specific objectives like classification, regression, or object detection.
AutoML is best for standard business problems with 1,000 to 1,000,000 samples; custom training is needed for novel architectures or extremely large or small datasets.
Deployed AutoML models require ongoing monitoring and retraining to handle concept drift, which can be automated with Vertex AI Pipelines.
These come up on the exam all the time. Here's how to tell them apart.
AutoML
Requires no coding – fully automated through console or API
Best for standard tasks like classification, regression, and forecasting
Limited control over model architecture and training details
Custom Training
Requires writing code in Python using TensorFlow, PyTorch, etc.
Needed for novel architectures, research, or unsupported data types
Full control over every hyperparameter, layer, and training loop
Low-Code Tools (Vertex AI Pipelines)
Visual drag-and-drop interface to build ML workflows
Generates underlying code but abstracts details from the user
Faster for standard pipelines but limited custom logic
Full-Code Development (Python SDK)
Everything written manually in Python using the Vertex AI SDK
Maximum flexibility for complex, custom logic and branching
Slower to develop but enables deep integration and automation
AutoML for Tabular Data
Input is structured data in rows and columns (CSV, BigQuery)
Supports regression, classification, and time-series forecasting
Feature importance explanations are based on column values
AutoML for Image Data
Input is image files (JPEG, PNG) stored in Cloud Storage
Supports classification, object detection, and segmentation
Explanations use saliency maps showing which pixels influenced the result
Mistake
AutoML means you don't need to prepare or clean your data at all – the tool does everything automatically.
Correct
AutoML requires clean, labelled data. It automates model selection, hyperparameter tuning, and evaluation, but data preparation – handling missing values, removing duplicates, correcting mislabelled data – is still your responsibility. Garbage in, garbage out applies.
The name 'AutoML' sounds like fully automatic, so beginners assume no manual work is required. In reality, data quality is the most important factor in model performance, and AutoML cannot fix messy data.
Mistake
AutoML always produces a model that is better than what a human data scientist could build.
Correct
AutoML is great for standard problems, but for cutting-edge tasks, novel architectures, or problems with very small or very large datasets, a skilled data scientist can often build a better model with custom code. AutoML is a productivity tool, not always the best option.
Marketing materials hype AutoML as 'expert-level performance', so beginners think it always outperforms humans. In practice, for unusual problems, custom work is superior.
Mistake
Low-code ML tools like Vertex AI Pipelines do not generate code at all – they are completely code-free.
Correct
Low-code tools reduce the amount of code you need to write, but they still generate underlying code (Python, YAML, etc.) that you can inspect and modify. 'Low-code' means less code, not zero code. Some configuration and occasional scripting are still required.
The term 'low-code' is often misinterpreted as 'no-code'. Beginners expect a completely visual experience, but real-world pipelines often require small code snippets for custom logic.
Mistake
Once an AutoML model is deployed, you never need to retrain or update it.
Correct
Models need periodic retraining as new data comes in to avoid concept drift – when the underlying patterns change. Vertex AI provides managed pipelines to automate this retraining on a schedule or trigger.
Beginners think machine learning models are like software that works forever. In reality, data distributions change over time, and models must be updated to stay accurate.
Reveal each answer, then mark whether you got it right. Score 60%+ to unlock the next chapter.
No, AutoML can be used entirely through the Google Cloud console web interface with no coding. However, if you want to automate steps or use advanced features, basic Python or the gcloud command-line tool is helpful.
Yes, Vertex AI AutoML has a text classification objective that includes sentiment analysis. You provide labelled text examples, and AutoML trains a model to predict sentiment categories (e.g., positive, negative, neutral).
AutoML automatically handles algorithm selection, hyperparameter tuning, and evaluation with no code. Custom training requires you to write your own training code (e.g., in TensorFlow or PyTorch) and gives you full control over the model architecture.
AutoML works best with at least 1,000 labelled examples per category, but it can produce reasonable results with as few as 100. With fewer than 100 examples per category, custom training with transfer learning may be more effective.
Yes, Vertex AI offers AutoML Edge, which trains models optimised for on-device inference. You can export the model in TensorFlow Lite format and integrate it into mobile or IoT applications.
A Vertex AI Pipeline is a low-code tool that lets you orchestrate ML workflows by connecting steps visually. You can create a pipeline that automatically runs AutoML training, evaluates the model, and deploys it, enabling repeatable MLOps.
You've finished AutoML and Low-Code ML with Vertex AI. Continue through the PMLE study guide to build a complete picture of the exam.
Done with this chapter?