Why this level matters
Task statement 1.1 of the AIF-C01 exam guide expects you to describe the types of data in AI models, the types of AI/ML learning, and the types of inferencing. Scenario questions describe a dataset or workload in plain language and expect you to name the category. This level gives you a repeatable classification procedure for each of the three.
Task 1: Classify the data
Ask two separate questions about any dataset. They are independent of each other.
Question A: is it labeled?
- Labeled data: each example comes with the correct answer (the label or target). Example: 50,000 emails, each tagged "spam" or "not spam".
- Unlabeled data: examples only, with no answer attached. Example: 50,000 emails with no tags.
Question B: what shape is it?
- Structured data: follows a fixed schema of rows and columns, as in a relational database table.
- Unstructured data: has no predefined schema. Images, free text, audio and video are the standard examples.
Within those two shapes, the exam guide names four specific types:
| Type | What it looks like | Example |
|---|---|---|
| Tabular | Rows and columns (structured) | Customer table: age, plan, monthly spend |
| Time-series | Values recorded in time order, each with a timestamp | Hourly CPU utilization, daily sales |
| Image | Pixel data (unstructured) | X-ray scans, product photos |
| Text | Natural language (unstructured) | Support tickets, reviews, contracts |
Do this when you read a scenario:
- Look for an answer column or tag. If one exists, the data is labeled.
- Look for a schema. Rows and columns mean structured or tabular; files of pixels or prose mean unstructured.
- Look for timestamps where order matters. If shuffling the rows would destroy the meaning, it is time-series.
Note that labeling and shape combine freely. A folder of X-rays tagged "fracture" or "no fracture" is labeled and unstructured image data.
Task 2: Choose the learning type
| Learning type | Data it needs | What it learns | Typical scenario |
|---|---|---|---|
| Supervised learning | Labeled | A mapping from inputs to known answers | Predict churn (yes/no) from past customers whose outcome is known |
| Unsupervised learning | Unlabeled | Patterns or groupings nobody defined in advance | Group customers into segments with no predefined categories |
| Reinforcement learning | No fixed dataset; an agent acts in an environment | A strategy that maximises cumulative reward through trial and error | A warehouse robot learning the fastest route |
Use this decision procedure:
- Does the scenario include historical examples with known outcomes, and the goal is to predict that outcome for new cases? Choose supervised.
- Is there data but no known answers, and the goal is to discover structure? Choose unsupervised.
- Is something taking a sequence of actions, receiving rewards or penalties, and improving over time? Choose reinforcement.
Then eliminate the other two explicitly. For the churn example: unsupervised does not fit because the labels exist and a specific outcome must be predicted; reinforcement does not fit because there is no agent taking actions for rewards.
Analogy: supervised learning is studying with an answer key, unsupervised learning is sorting a pile of documents into folders with no instructions, and reinforcement learning is tuning an Auto Scaling policy by trying settings and watching the bill.
Task 3: Choose the inferencing type
Level 1 defined inferencing as using a trained model to produce predictions. The exam guide lists four types. Amazon SageMaker AI offers each as an inference option, and its documentation gives the figures below.
| Type | Use it when | Documented limits in SageMaker AI |
|---|---|---|
| Real-time inference | You need an immediate, low-latency response and traffic is sustained. Uses a persistent endpoint. | Payloads up to 25 MB; 60 seconds processing for regular responses |
| Serverless inference | Traffic is intermittent or unpredictable, with idle periods, and the workload can tolerate cold starts. No instances to manage; you do not pay for idle time. | Payloads up to 4 MB; processing up to 60 seconds |
| Asynchronous inference | Payloads are large or processing is long, and requests can wait in a queue. The endpoint can scale down to zero when there are no requests. | Payloads up to 1 GB; processing up to one hour |
| Batch inference (Batch Transform) | A large dataset is available up front, nobody is waiting for a response, and no persistent endpoint is needed. | Datasets that are gigabytes in size; processing times of days |
A cold start is the delay while a serverless endpoint launches compute after a period with no traffic.
Apply these steps in order:
- Is anyone waiting for the answer? If no, and all the data is already collected, choose batch. Example: score all 2 million customers overnight for next week's campaign.
- Is the payload large or the processing long? If one request is hundreds of megabytes or takes minutes, choose asynchronous. Example: analyse a 600 MB video upload and notify the user when done.
- Is traffic sporadic with idle gaps? If yes, and an occasional slow first response is acceptable, choose serverless. Example: an internal tool used a few times a day.
- Otherwise, with steady traffic and a strict low-latency requirement, choose real-time. Example: fraud check during card checkout.
Common confusion: asynchronous is not batch. Asynchronous inference handles individual requests as they arrive and queues them. Batch processes a whole dataset in one offline job.
Tip (practitioner advice, not from the exam guide): underline three things in every inferencing scenario: payload size, acceptable latency, and traffic pattern. Those three facts nearly always decide the answer.
What the exam asks about this
- A dataset description, and you classify it (labeled or unlabeled; structured, unstructured, tabular, time-series, image or text).
- A business scenario, and you select supervised, unsupervised or reinforcement learning. Distractors are usually the other two types.
- A workload description with latency, volume, payload or traffic clues, and you select batch, real-time, asynchronous or serverless inferencing.
- You are not expected to write code or configure endpoints; you are expected to recognise the right category.
Key takeaways
- Labeled data has known answers attached; unlabeled data does not. This is separate from whether data is structured or unstructured.
- Tabular data is rows and columns; time-series data is ordered by time; image and text are unstructured.
- Supervised needs labels, unsupervised finds patterns without labels, reinforcement learns from rewards through trial and error.
- Real-time: low latency, sustained traffic. Serverless: intermittent traffic, cold starts tolerated. Asynchronous: large payloads (up to 1 GB) or long processing (up to one hour), queued. Batch: whole dataset offline, no persistent endpoint.
- Decide inferencing from payload size, latency and traffic pattern.