Practical

These are the ten questions from the AI Security Analyst Orientation in TryHackMe’s AI Security Threats room, supplied by Ramesh. They appear below in their original question order, with text answers and explanations.

Questions 6 and 7 use simplified wording. For those two, expected room answer means the choice suggested by the room’s teaching. The explanation also gives the more precise meaning. The supplied question images do not show which individual answers the quiz accepted.

1. Machine Learning is a subfield of which broader field?

Correct answer: C — Artificial Intelligence.

Why: Artificial intelligence, or AI, is the broad field. It includes computer systems that do tasks such as recognizing images, making predictions, or working with language. Machine learning, or ML, is one part of AI. It learns patterns from data.

For example, an email filter can learn from examples of unwanted messages. That is machine learning, so it is also part of AI.

Remember: Machine learning belongs inside artificial intelligence. Deep learning is a smaller part of machine learning, not the broader field. Google’s definitions of AI and ML.

2. A model performs perfectly on its training data but poorly on data it has never seen before. What is this called?

Correct answer: B — Overfitting.

Why: The model has learned the training examples too closely, including details that do not help with new examples. It does well on familiar data but struggles when the data changes.

Imagine remembering the answers to ten practice questions without learning how to solve them. You can repeat those answers, but a new question is difficult.

Training data is the data used for learning. Separate test data helps check how well the model handles new examples. Underfitting is different: the model has not learned enough useful patterns and usually performs poorly even on training data. Google’s explanation of overfitting.

3. A security tool analyses network traffic without any pre-labelled examples and flags unusual patterns on its own. Which type of ML is this?

Correct answer: C — Unsupervised learning.

Why: The key words are without any pre-labelled examples. A label is a known answer attached to an example, such as “normal” or “suspicious.” Here, the tool looks for patterns without being given those answers first.

It might notice that one computer suddenly sends much more information than similar computers. That difference may deserve investigation. It does not prove an attack: a backup or a software update might also explain it.

Remember: In this question, no labels are supplied, so unsupervised learning is the matching choice. Google’s introduction to unsupervised learning.

4. A spam filter is trained using emails already labelled as “spam” or “not spam”. Which ML algorithm type does this describe?

Correct answer: D — Supervised learning.

Why: Each training email comes with a known answer: “spam” or “not spam.” The model learns the relationship between the email and its label. It then uses those patterns to judge new emails.

This is also classification, which means choosing a category. The categories here are spam and not spam.

Remember: Training examples with known answers indicate supervised learning. “Supervised” does not mean a person must watch every prediction after training. Google’s explanation of supervised learning.

5. Which layer in a neural network produces the final prediction or classification?

Correct answer: D — Output layer.

Why: A neural network passes information through layers. The input layer receives the information. Hidden layers process it. The output layer produces the model’s final values.

For a spam filter, the final value might be a score used to choose between “spam” and “not spam.” The application may turn that score into the displayed label.

Remember: Input receives → hidden layers process → output gives the result. A weight is a learned number used in these calculations; “weight layer” is not the correct choice here. Google’s guide to neural-network layers.

6. What is the key distinction between Machine Learning and Deep Learning?

Expected room answer: C — Deep Learning does not require labelled data and can self-learn from raw input.

Why this matches the room: The lesson points to deep learning’s ability to learn useful patterns from raw data. For example, a language model can learn to predict missing text using the surrounding text. A person does not need to label every sentence manually. TryHackMe’s Building Blocks of AI lesson.

The more accurate explanation: Deep learning is a type of machine learning that uses neural networks with many processing layers. It can use labelled data or learn without manually supplied labels. Other machine learning methods can also use unlabelled data.

So labels are not the main difference. The network’s layers are. Deep learning is not always faster, and it can work with images, sound, and other data as well as text. Deep Learning textbook: representation learning.

7. During LLM pre-training, which algorithm adjusts the model’s parameters after each incorrect prediction?

Expected room answer: D — Backpropagation.

Why this matches the room: The TryHackMe lesson uses “backpropagation” when describing how training adjusts the model after prediction errors.

The more accurate explanation: Parameters are the model’s adjustable numbers, including its weights. During training, the model makes predictions and the training system measures the error. Backpropagation works backward through the network to calculate how each parameter affects that error.

An optimizer then uses those calculations to change the parameters. A — Gradient descent names an update method more precisely. Backpropagation calculates the information for the change; the optimizer applies the change. PyTorch’s training-loop explanation.

Remember: Backpropagation calculates → optimizer updates. Training often groups examples into batches, so it need not update once after every individual wrong prediction.

8. What was the key innovation introduced by transformer neural networks in Google’s 2017 “Attention is All You Need” paper?

Correct answer: C — The attention mechanism — assigning importance to words based on context.

Why: Attention helps a model use relevant information from different parts of the text. For example, in “The animal was tired, so it slept,” nearby information helps connect “it” with “the animal.”

More precisely, the model processes tokens, which are small pieces of text. A token can be a word, part of a word, or punctuation. Attention combines information from these pieces according to the context.

The paper’s main contribution was a new network design built around attention. Attention itself had already been studied before 2017. Original Transformer paper, earlier attention research.

9. What does RLHF stand for?

Correct answer: B — Reinforcement Learning from Human Feedback.

Why: RLHF uses people’s judgments to help guide a model’s behavior. In one common approach, people compare different answers and choose which they prefer. Those choices help train a reward model, which gives scores used in further training.

For example, reviewers might prefer an answer that explains a term clearly over an answer that is confusing. The training process encourages responses that receive better scores.

Human feedback does not guarantee that every later answer is true or safe. RLHF is one training approach, and not every language model uses it. Research on training language models with human feedback.

10. Which correctly represents the AI hierarchy from broadest to most specific?

Correct answer: D — AI → ML → DL → LLMs.

Why: Read this as groups inside larger groups:

• Artificial Intelligence (AI): the broad field.
• Machine Learning (ML): a part of AI that learns patterns from data.
• Deep Learning (DL): a part of ML that uses neural networks with many layers.
• Large Language Models (LLMs): modern language models built using deep learning.

The arrows show which group contains the next group. They do not show steps that every AI system must follow. Not all AI uses ML, and not all deep learning is about language. Google’s glossary of AI, ML, deep models, and LLMs.

Quick revision: the ten answers in order

1. C — Artificial Intelligence.
2. B — Overfitting.
3. C — Unsupervised learning.
4. D — Supervised learning.
5. D — Output layer.
6. C — Expected room answer. Read the correction about labels above.
7. D — Expected room answer: Backpropagation. Read the distinction from A — Gradient descent above.
8. C — The attention mechanism.
9. B — Reinforcement Learning from Human Feedback.
10. D — AI → ML → DL → LLMs.