Language Models

Question 1

What is an embedding?

Answer 1

It is a low-dimensional representation of sparse data.

Question 2

For what kind of data is an embedding specifically useful? Example?

Answer 2

For sparse data – such as a one-hot encoding.

Question 3

What is word2vec?

Answer 3

An embedding for English-language words.

Question 4

How do you create an embedding?

Answer 4

You use a neural network, where the prediction task is based on similarity. Then the first hidden layer – the neurons the input values are mapped to – is called the “projection layer” and projects the input into a space that is most useful for solving that task.

Question 5

What is the projection layer?

Answer 5

It’s the first hidden layer in a neural network, because it’s taking the input values and projecting them onto an embedding space.

Question 6

What is an example of a task you would use to train a neural network to create an embedding?

Answer 6

Get words that frequently appear in the same context.

Question 7

How is embedding related to generalization?

Answer 7

The embedding space captures similar concepts, relationships, and differences between examples. This is what allows the model to generalize outside of the training set, because it can reason about new examples based on what they are conceptually, and what their relationships to known examples are.

Question 8

What, specifically, is a language model?

Answer 8

It’s a model that predicts the probability of a token, or sequence, occurring within a longer sequence.

Question 9

What is a representation model?

Answer 9

It’s a model that learns useful vector representations of text.

Question 10

How do “language model” and “representation model”, as concepts, related to transformers?

Answer 10

A representation model is for encoding – so the encoder part is one. A language model is for generating – so the decoder part is one.

Question 11

What is an N-gram? What is a bi-gram?

Answer 11

A sequence of N words. For a bigram, N=2.

Question 12

What does the output of a language model look like? How does the application use it?

Answer 12

A distribution over possible words or phrases. For example, 30% cat 70% tiger. The application usually picks probabilistically from the words above a certain threshold.

Question 13

How are N-grams related to context? What is the tradeoff?

Answer 13

The bigger N is, the more context you have. However, this reduces the number of times in your dataset you would see each N-gram.

Question 14

What is the transformer?

Answer 14

It is the architecture used for LLM applications. It is composed of an encoder and a decoder, and relies on attention for context.

Question 15

What is an encoder?

Answer 15

The encoder processes input text into some intermediate representation.

Question 16

What is a decoder?

Answer 16

The decoder converts an intermediate representation into output text.

Question 17

When would you want to use an encoder-only architecture?

Answer 17

If you only want the embedding.

Question 18

When would you want to use a decoder-only architecture?

Answer 18

If you only care about generating new tokens.

Question 19

How do transformers solve the context problem?

Answer 19

They use a self-attention layer.

Question 20

What does BERT stand for?

Answer 20

Bidirectional Encoder Representations from Transformers.

Question 21

What is the name of the really famous transformers paper?

Answer 21

“Attention is all you need.”

Question 22

BERT stands for Bidirectional Encoder Representations from Transformers. What does “bidirectional” mean?

Answer 22

It means the self-attention looks at both the preceding and the following tokens.

Question 23

Is BERT a representation model or a language model?

Answer 23

It’s a representation model, since its purpose is to give good encodings of words.

Question 24

How do you avoid “cheating” with a bidirectional self-attention layer?

Answer 24

You train using masked language modeling.

Question 25

What is masked language modeling?

Answer 25

It’s a form of unsupervised learning where you take complete sequences of words and then “mask” (or erase) certain tokens to create training examples – the model then tries to predict the missing tokens.

Question 26

What does the “foundational LLM” refer to?

Answer 26

A foundational LLM would be a basic one that has yet to adapt to your application’s specific needs.

Question 27

What is fine-tuning?

Answer 27

It’s when you take your trained LLM and do follow-up training on examples specific to your task.

Question 28

What is the cost of fine-tuning in terms of training data?

Answer 28

Usually small, you only need a few thousand examples.

Question 29

What is the cost of fine-tuning in terms of performance?

Answer 29

It is quite expensive to fine-tune because you have to do backpropagation on the full set of parameters, which could be in the billions.

Question 30

What is parameter-efficient tuning?

Answer 30

It’s when you only adjust a subset of parameters per-iteration.

Question 31

What is distillation?

Answer 31

It’s when you create a smaller version of an LLM.

Question 32

Why would you want to do distillation?

Answer 32

Because the smaller “distilled” model will be faster and more efficient.

Question 33

What is the most common way to do distillation?

Answer 33

Using a teacher-student relationship where the student model is the distilled model.

Question 34

What is the difference between one-shot and few-shot?

Answer 34

These terms are used in the context of prompt engineering, where you are providing examples to the LLM in the prompt to teach it what kind of response you want. One-shot would mean you only give one example. Few-shot would mean many examples.

Question 35

What is the difference between online and offline inference?

Answer 35

Online means you have the LLM respond at serving time. Offline means you’re caching predictions made in bulk in advance.

Question 36

What is a “sparse encoding”?

Answer 36

It’s when we try to regularize an encoding layer by encouraging the embedding to be just a few dimensions, and 0 for the rest.

Question 37

What are “basis functions”?

Answer 37

They are the functions that are applied to the input data to get its embedding.

Question 38

How do basis functions relate to sparse encoding?

Answer 38

Sparse encoding would be trying to reduce the embedding to a weighted combination of just a few of the basis functions, and regularize out the rest.

Question 39

What is an autoencoder?

Answer 39

An autoencoder is an architecture focused on finding an embedding. It has an encoding layer and a decoding layer.

Question 40

What is a “latent space”?

Answer 40

It’s a lower-dimensional representation of your data, that still captures essential meaning and patterns.

Question 41

How are autoencoders and transformers different?

Answer 41

Autoencoders are designed specifically to find an embedding for individual examples, and don’t try to model relationships. That is because their purpose is to find a low-dimensionality latent space that can be used to compress data. Transformers are designed for sequential data, which is why they include the attention mechanism that autoencoders lack.

Question 42

What does VAE stand for?

Answer 42

Variational AutoEncoder.

Question 43

What is a variational autoencoder?

Answer 43

It’s an autoencoder that learns a probabilistic model rather than exact outputs.

Question 44

Why would you want to use a variational autoencoder?

Answer 44

Because you could use it for data generation. Take your original examples, encode them, and then sample from the resulting distribution to get new examples with a similar encoding – which will be similar to your original unencoded example, if the embedding is good.

Question 45

How do you train an autoencoder?

Answer 45

You take the input, reduce it to the latent space in the encoder, and then decode it from that latent space. So basically, take 100 dimensions down to 2, then expand back to 100. The loss function compares the output, which is a reconstruction of the input, to the real input. The better the latent space is at giving the decoder the information it needs, the lower the loss will be.

Question 46

How are autoencoders related to EBR?

Answer 46

You use an autoencoder to map points to the latent space, and then search by similarity.

Question 47

What is a denoising autoencoder and why would you want to use it?

Answer 47

A denoising autoencoder adds noise to the input data during training, and then compares the output to the non-noisy data. So it learns to deal with noise. You’d use this when real-world data is noisy.

Question 48

What is Softmax?

Answer 48

It’s when you reduce a set of numbers such that they add up to 1.

Question 49

Why would you want to use Softmax, in general?

Answer 49

You would use it when each value in the set is supposed to represent a probability. In that case you need all the probabilities to add up to 1.

Question 50

Why would you use Softmax, in deep learning?

Answer 50

You would use it in the output layer, when you are outputting a probability, such as predicting what word will come next.

Question 51

What is the temperature parameter in Softmax and what happens when you change it?

Answer 51

The temperature parameter controls how flat the distribution is – higher T means more chances are given to less-popular options, making the probability distribution flatter and thus less confident and more random.

Question 52

What is an embedding matrix?

Answer 52

It’s what you turn your vocabulary into after you learn the embedding.

Question 53

What does it mean to say that you can do math on embeddings?

Answer 53

It means that since each direction in the embedding space represents a concept, you can find add that concept to a word to get a new word, where you find the concept based on a diff between other words. Like “sushi” + (“germany - japan”) = “bratwurst” (switching from germany to japan concept) or “cat” + (“dogs” - “dog”) = “cats” (adding plurality concept).

Question 54

What do you use to measure vector similarity?

Answer 54

Dot product.

Question 55

What is an attention head trying to do?

Answer 55

It is trying to let each word ask some question of each other word, and then conceptually update itself based on that other word if the answer is “yes.”

Question 56

What are the three parts of an attention head and what do they do?

Answer 56

Query matrix – lets each word represent itself as some conceptual “question”; Key matrix – lets each word represent itself as some conceptual “answer”; Value matrix – lets each word represent itself as some conceptual “update” to apply to another word.

Question 57

What is the query vector in attention?

Answer 57

It’s a representation of a question that an embedded vector asks – like “are there any adjectives in front of me?”

Question 58

What is the query matrix in attention?

Answer 58

It’s a matrix you multiply your embedded vector by to get a question specific to that (embedded) word/phrase/concept.

Question 59

What is the key vector, in attention?

Answer 59

It’s a representation of an answer that an embedded vector provides for a particular question, like “I am an adjective.”

Question 60

What is the key matrix in attention?

Answer 60

It’s a matrix you multiple your embedded vector by to get an answer specific to that (embedded) word/phrase/concept.

Question 61

What is the value vector in attention?

Answer 61

For some embedded vector, it’s the “update” that that embedded vector provides, based on the question being asked. Like if the question is about adjectives, it might be “here is my adjective concept, you should add this to yourself.”

Question 62

How do you know if you should apply an update in attention?

Answer 62

You take the dot product of the question (query for vector A) and the answer (key for vector B) and if they’re similar, you apply B’s update (value for vector B) to A.

Question 63

Explain what an “attention head” is and what “multi-headed attention” means

Answer 63

An attention head is the set of attention operations applied on one set of query/key/value matrices, to let all the words ask each other one question. “Multi-headed attention” means you actually have a bunch of different attention heads, so you’re capturing a bunch of different concepts with this question/answer scheme.

Question 64

Distinguish between “many attention layers” and “multi-headed attention”

Answer 64

An attention layer is one set of multi-headed attention operations – like 100 different attention heads running in parallel. Because they’re running in parallel, all the questions are independent of each other. If you want concepts to be related and learn from each other, you want multiple attention layers sequentially, which is of course more expensive.

Question 65

Distinguish between “encoder” and “projector”

Answer 65

Projector means a layer that maps data from one space to another. Encoder means mapping data to a different representation, which captures structures and patterns. An encoder is technically a class of projector, but it’s so specific and unique that we delineate it with its own term.