Lecture 3

Question 1

Waarvoor gebruiken streamingdiensten de kijkgedragingen van gebruikers?

Answer 1

Ze vergelijken jouw kijkgedrag met dat van andere gebruikers om te voorspellen welke shows je leuk vindt.

Question 2

Welke wiskundige methode ligt ten grondslag aan het aanbevelingssysteem bij streamingdiensten?

Answer 2

Lineaire logistische regressie.

Question 3

Hoe werkt lineaire logistische regressie in het context van aanbevelingen?

Answer 3

Het model voorspelt de waarschijnlijkheid dat je een bepaalde show leuk vindt op basis van gelijkenissen in kijkgedrag.

Question 4

Wat is het verschil tussen sceptici en enthousiastelingen als het gaat om AI?

Answer 4

Sceptici focussen vaak op de basiswiskunde en bouwstenen van AI, terwijl enthousiastelingen zich vooral richten op de coole eindresultaten van AI.

Question 5

Wat is het mechanisme achter veel van de coole dingen die AI kan doen?

Answer 5

Het analyseren van data (zoals gebruikersgedrag) en voorspellen op basis van statistische modellen zoals lineaire logistische regressie.

Question 6

What is a limitation of the linear/logistic regression model in prediction?

Answer 6

It struggles with situations where the predictor needs to capture “either/or” logic, like deciding if a movie is good based on whether either Meryl Streep or The Rock stars in it.

Question 7

Give an example illustrating the limitation of logistic regression with the “either/or” problem.

Answer 7

If a movie has Meryl Streep or The Rock, you predict it’s good. But if it has both, logistic regression struggles because they excel in different genres, so combining them confuses the model.

Question 8

What is the “XOR problem” in AI?

Answer 8

It’s a classic problem where AI models like logistic regression cannot learn the logical function of “this or that” (exclusive OR).

Question 9

Why do we need neural networks to solve the XOR problem?

Answer 9

Because neural networks can perform multiple logistic regressions in parallel and combine their results to capture complex logical relationships.

Question 10

How do neural networks improve on simple logistic regression?

Answer 10

Neural networks build intermediate predictors by combining logistic regressions in layers, which extract useful combinations of raw data and improve understanding.

Question 11

What is the role of intermediate predictors in neural networks?

Answer 11

They create more informative features from the original data, enabling the AI to make more accurate predictions by considering more complex patterns.

Question 12

What is the “Universal Approximation Theorem”?

Answer 12

It states that neural networks can approximate any relationship between inputs and outputs, meaning they can model almost any logical or functional pattern.

Question 13

Why can’t simple logistic regression handle XOR but neural networks can?

Answer 13

Because logistic regression only models linear relationships, while neural networks combine many logistic regressions in layers, allowing them to model non-linear, complex relationships like XOR.

Question 14

Summarize why neural networks are a much better way to predict data compared to simple logistic regression

Answer 14

Neural networks create multiple intermediate steps that transform raw data into meaningful features, enabling them to capture complex patterns and logic that logistic regression cannot

Question 15

Does ChatGPT work exactly like a neural network?

Answer 15

Not exactly. It uses neural networks as building blocks within a broader architecture called a transformer model

Question 16

What are language models like ChatGPT trained to do?

Answer 16

Predict the next word in a sequence—not the correct answer, just the most likely one based on previous words.

Question 17

Why is predicting the next word difficult for AI?

Answer 17

While it’s a basic task, doing it well requires a lot of background knowledge, context awareness, and handling of word order.

Question 18

Why don’t regular neural networks work well for language tasks?

Answer 18

Because they don’t consider word order, which is crucial in understanding language.

Question 19

What model solves this problem of word order and meaning?

Answer 19

The transformer model.

Question 20

What does the encoder part of the transformer do?

Answer 20

Turns words into numbers that best capture the meaning of the input.

Question 21

What are embeddings?

Answer 21

Numerical representations of words based on their co-occurrence with other words.

Question 22

How are embeddings created in basic neural networks?

Answer 22

By trying to predict the next word and observing which words occur together frequently.

Question 23

What is positional encoding in transformers?

Answer 23

A numerical representation added to embeddings to capture a word’s position in the sentence.

Question 24

What happens when you combine embedding and positional encoding?

Answer 24

You get a richer numerical representation that includes both meaning and position (e.g., won = {0.4, 0.4, 0} + {1, 0.9, -0.8} = {1.4, 1.3, 0.8}).

Question 25

What are skip connections in transformers?

Answer 25

Shortcuts that copy original input numbers to later layers to retain information and stabilize it (with “Add & Norm”).

Question 26

What does FNN (Feed Forward Network) do in a transformer?

Answer 26

It transforms numbers into more informative numbers—just like what neural networks do.

Question 27

What is attention in language models?

Answer 27

It allows the model to focus on relevant surrounding words to improve a word’s representation.

Question 28

What is multi-head attention?

Answer 28

The model looks at the word from multiple angles, each focusing on different contextual relationships.

Question 29

What are the Query, Key, and Value in attention?

Answer 29

- Query: What you're looking for (like a question) - Key: What each word offers (the content) - Value: The actual information retrieved

Question 30

How does the attention mechanism work for a word like “won”?

Answer 30

The model calculates similarity between the Query of “won” and Keys of other words (like “prize”), then uses the corresponding Values to adjust the representation.

Question 31

What does the equation "Won = 0.6 x 'won' + 0.3 x 'prize' + 0.02 x 'home'" represent?

Answer 31

A weighted combination of word meanings based on attention scores that help define “won” more accurately in context.

Question 32

What does the decoder do in a transformer model?

Answer 32

Takes the final numerical representations and predicts the next word.

Question 33

What part of the transformer does GPT use?

Answer 33

Only the decoder part, since its job is to generate text.

Question 34

What kind of machine learning problems only use the encoder part?

Answer 34

Tasks like weather forecasting, where you don’t need to output text.

Question 35

How is GPT trained beyond regular prediction tasks?

Answer 35

It’s updated using human feedback through reinforcement learning (e.g., RLHF—Reinforcement Learning from Human Feedback).

Question 36

Why is GPT not “just” a language model anymore?

Answer 36

It can also function as a calculator, search engine, reasoning assistant, and more, by integrating multiple capabilities.