Ask HN: Can someone ELI5 transformers and the “Attention is all we need” paper?

Okay, here's my attempt!

First, we take a sequence of words and represent it as a grid of numbers: each column of the grid is a separate word, and each row of the grid is a measurement of some property of that word. Words with similar meanings are likely to have similar numerical values on a row-by-row basis.

(During the training process, we create a dictionary of all possible words, with a column of numbers for each of those words. More on this later!)

This grid is called the "context". Typical systems will have a context that spans several thousand columns and several thousand rows. Right now, context length (column count) is rapidly expanding (1k to 2k to 8k to 32k to 100k+!!) while the dimensionality of each word in the dictionary (row count) is pretty static at around 4k to 8k...

Anyhow, the Transformer architecture takes that grid and passes it through a multi-layer transformation algorithm. The functionality of each layer is identical: receive the grid of numbers as input, then perform a mathematical transformation on the grid of numbers, and pass it along to the next layer.

Most systems these days have around 64 or 96 layers.

After the grid of numbers has passed through all the layers, we can use it to generate a new column of numbers that predicts the properties of some word that would maximize the coherence of the sequence if we add it to the end of the grid. We take that new column of numbers and comb through our dictionary to find the actual word that most-closely matches the properties we're looking for.

That word is the winner! We add it to the sequence as a new column, remove the first-column, and run the whole process again! That's how we generate long text-completions on word at a time :D

So the interesting bits are located within that stack of layers. This is why it's called "deep learning".

The mathematical transformation in each layer is called "self-attention", and it involves a lot of matrix multiplications and dot-product calculations with a learned set of "Query, Key and Value" matrixes.

It can be hard to understand what these layers are doing linguistically, but we can use image-processing and computer-vision as a good metaphor, since images are also grids of numbers, and we've all seen how photo-filters can transform that entire grid in lots of useful ways...

You can think of each layer in the transformer as being like a "mask" or "filter" that selects various interesting features from the grid, and then tweaks the image with respect to those masks and filters.

In image processing, you might apply a color-channel mask (chroma key) to select all the green pixels in the background, so that you can erase the background and replace it with other footage. Or you might apply a "gaussian blur" that mixes each pixel with its nearest neighbors, to create a blurring effect. Or you might do the inverse of a gaussian blur, to create a "sharpening" operation that helps you find edges...

But the basic idea is that you have a library of operations that you can apply to a grid of pixels, in order to transform the image (or part of the image) for a desired effect. And you can stack these transforms to create arbitrarily-complex effects.

The same thing is true in a linguistic transformer, where a text sequence is modeled as a matrix.

The language-model has a library of "Query, Key and Value" matrixes (which were learned during training) that are roughly analogous to the "Masks and Filters" we use on images.

Each layer in the Transformer architecture attempts to identify some features of the incoming linguistic data, an then having identified those features, it can subtract those features from the matrix, so that the next layer sees only the transformation, rather than the original.

We don't know exactly what each of these layers is doing in a linguistic model, but we can imagine it's probably doing things like: performing part-of-speech identification (in this context, is the word "ring" a noun or a verb?), reference resolution (who does the word "he" refer to in this sentence?), etc, etc.

And the "dot-product" calculations in each attention layer are there to make each word "entangled" with its neighbors, so that we can discover all the ways that each word is connected to all the other words in its context.

So... that's how we generate word-predictions (aka "inference") at runtime!

By why does it work?

To understand why it's so effective, you have to understand a bit about the training process.

The flow of data during inference always flows in the same direction. It's called a "feed-forward" network.

But during training, there's another step called "back-propagation".

For each document in our training corpus, we go through all the steps I described above, passing each word into our feed-forward neural network and making word-predictions. We start out with a completely randomized set of QKV matrixes, so the results are often really bad!

During training, when we make a prediction, we KNOW what word is supposed to come next. And we have a numerical representation of each word (4096 numbers in a column!) so we can measure the error between our predictions and the actual next word. Those "error" measurements are also represented as columns of 4096 numbers (because we measure the error in every dimension).

So we take that error vector and pass it backward through the whole system! Each layer needs to take the back-propagated error matrix and perform tiny adjustments to its Query, Key, and Value matrixes. Having compensated for those errors, it reverses its calculations based on the new QKV, and passes the resultant matrix backward to the previous layer. So we make tiny corrections on all 96 layers, and eventually to the word-vectors in the dictionary itself!

Like I said earlier, we don't know exactly what those layers are doing. But we know that they're performing a hierarchical decomposition of concepts.

Hope that helps!

The Yannic kilcher review is quite good.

https://youtu.be/iDulhoQ2pro

I can't ELI5 but I can ELI-junior-dev. Tl;dw:

Transformers work by basically being a differentiable lookup/hash table. First your input is tokenized and (N) tokens (this constitutes the attention frame) are encoded both based on token identity and position in the attention frame.

Then there is an NxN matrix that is applied to your attention frame "performing the lookup query" over all other tokens in the attention frame, so every token gets a "contextual semantic understanding" that takes in both all the other stuff in the attention frame and it's relative position.

Gpt is impressive because the N is really huge and it has many layers. A big N means you can potentially access information farther away. Each layer gives more opportunities to summarize and integrate long range information in a fractal process.

Two key takeaways:

- differentiable hash tables

- encoding relative position using periodic functions

NB: the attention frame tokens are actually K-vectors (so the frame is a KxN matrix) and the query matrix is an NxNxK tensor IIRC but it's easier to describe it this way

ELI5 is tricky as details have to be sacrificed, but I'll try.

An attention mechanism is when you want a neural network to learn the function of how much attention to allocate to each item in a sequence, to learn which items should be looked at.

Transformers is a self-attention mechanism, where you ask the neural network to 'transform' each element by looking at its potential combination with every other element and using this (learnable, trainable) attention function to decide which combination(s) to apply.

And it turns out that this very general mechanism, although compute-intensive (it considers everything linking with everything, so complexity quadratic to sequence length) and data-intensive (it has lots and lots of parameters, so needs huge amounts of data to be useful) can actually represent many of things we care about in a manner which can be trained with the deep learning algorithms we already had.

And, really, that's the two big things ML needs, a model structure where there exists some configuration of parameters which can actually represent the thing you want to calculate, and that this configuration can actually be determined from training data reasonably.

The Illustrated Transfomer ( https://jalammar.github.io/illustrated-transformer/ ) and Visualizing attention ( https://towardsdatascience.com/deconstructing-bert-part-2-vi... ), are both really good resources. For a more ELI5 approach this non-technical explainer ( https://www.parand.com/a-non-technical-explanation-of-chatgp... ) covers it at a high level.

Suppose someone asked you to complete the sentence:

“After I woke up and made breakfast, I drank a glass of …”

In America one might say the most likely next words are “orange juice”, or “apple juice” but not “sports car” which has nothing to do with the sentence.

Ultimately this is what language models do, given a sequence of data (in this case words) predict the most likely next word(s).

For attention, when you read the sentence, which words stood out as more important? Probably woke up, breakfast, and glass while the words after, I, and made were less important to completing the sentence.

That is, you paid more attention to the important words to understand how to complete the sentence.

The “attention mechanism” in language models is a way to let the models learn which words are important in sentences and pay more attention to them too when completing sentences, just like a person would do as in the example above.

Further, it turns out this attention mechanism lets the models do lots of interesting things even without other fancy model techniques. That is “attention is all you need”.

Those Computerphile videos[0] by Rob Miles helped me understand transformers. He specifically references the "Attention is all you need" paper.

And for a deeper dive, Andrej Kharpaty has this hands-on video[1] where he builds a transformer from scratch. You can check-out his other videos on NLP as well they are all excellent.

[0] https://youtu.be/rURRYI66E54, https://youtu.be/89A4jGvaaKk

[1] https://youtu.be/kCc8FmEb1nY

Well here is my (a bit cynical) take on it.

In the beginning, there was the matrix multiply. A simple neural network is a chain of matrix multiplies. Let's say you have your data A1 and weights W1 in a matrix. You produce A2 as A1xW1. Then you produce A3 as A2xW2, and so on. There are other operations in there like non-linearities (so that you can actually learn something interesting) and fancy batch norms, but let's forget about those for now. The problem with this is, it's not very expressive. Let's say your A1 matrix has just 2 values, and you want the output to be their product. Can you learn a weight matrix that performs multiplication of these inputs? No you can't. Multiplication must be simulated by piecing together piecewise linear functions. To perform multiplication, the weight matrix W would also need to be produced by the network. Transformers do basically that. In the product A*W you replace A with (AxW1), W with (AxW2), and multiply those together: (AxW1)x(AxW2) And then do it once more for good measure: (AxW1)x(AxW2)x(AxW3). Boom, Nobel prize. Now your network can multiply, not just add. OK it's actually a bit more complicated, there is for example a softmax in the middle to perform normalisation, which in general helps during numerical optimisation: softmax((AxW1)x(AxW2))x(AxW3). There are then fancy explanations that try to retrospectively justify this as a "differentiable lookup table" or somesuch nonsense, calling the 3 parts "key", "query" and "value", which help make your paper more popular. But the basic idea is not so complicated. A Transformer then uses this operation as a building block (running them in parallel an in sequence) to build giant networks that can do really cool things. Maybe you can teach networks to divide next and then you get the next Nobel prize.

It works like this:

First, convert the input text to a sequence of token numbers (2048 tokens with 50257 possible token values in GPT-3) by using a dictionary and for each token, create a vector with 1 at the token index and 0 elsewhere, transform it with a learned "embedding" matrix (50257x12288 in GPT-3) and sum it with a vector of sine and cosine functions with several different periodicities.

Then, for each layer, and each attention head (96 layers and 96 heads per layer in GPT-3), transform the input vector by query, key and value matrices (12288x128 in GPT-3) to obtain a query, key and value vector for each token. Then for each token, compute the dot product of its query vector with the key vectors of all previous tokens, scale by 1/sqrt of the vector dimension and normalize the results so they sum to 1 by using softmax (i.e. applying e^x and dividing by the sum), giving the attention coefficients; then, compute the attention head output by summing the value vectors of previous tokens weighted by the attention coefficients. Now, for each token, glue the outputs for all attention heads in the layer (each with its own key/query/value learned matrices), add the input and normalize (normalizing means that the vector values are biased and scaled so they have mean 0 and variance 1).

Next, for the feedforward layer, apply a learned matrix, add a learned vector and apply a ReLU (which is f(x) = x for positive x and f(x) = kx with k near 0 for negative x), and do that again (12288x49152 and 49152x12288 matrices in GPT-3, these actually account for around 70% of the parameters in GPT-3), then add the input before the feedforward layer and normalize.

Repeat the process for each layer, each with their own matrices, passing the output of the previous layer as input. Finally, apply the inverse of the initial embedding matrix and use softmax to get probabilities for the next token for each position. For training, train the network so that they are close to the actual next token in the text. For inference, output a next token according to the top K tokens in the probability distribution over a cutoff and repeat the whole thing to generate tokens until an end of text token is generated.

I'll throw my hat in the ring.

A transformer is a type of neural network that, like many networks before, is composed of two parts: the "encoder" that receives a text and builds an internal representation of what the text "means"[1], and the "decoder" that uses the internal representation built by the encoder to generate an output text. Let's say you want to translate the sentence "The train is arriving" to Spanish.

Both the encoder and decoder are built like Lego, with identical layers stacked on top of each other. The lowest lever of the encoder looks at the input text and identifies the role of individual words and how they interact with each other. This is passed to the layer above, which does the same but at a higher level. In our example it would be as if the first layer identified that "train" and "arrive" are important, then the second one identifies that "the train" and "is arriving" are core concepts, the third one links both concepts together, and so on.

All of these internal representations are then passed to the decoder (all of them, not just the last ones) which uses them to generate a single word, in this case "El". This word is then fed back to the decoder, that now needs to generate an appropriate continuation for "El", which in this case would be "tren". You repeat this procedure over and over until the transformer says "I'm done", hopefully having generated "El tren está llegando" in the process.

The attention mechanism already existed before transformers, typically coupled with an RNN. The key concept of the transformer was building an architecture that removed the RNN completely. The negative side is that it is a computationally inefficient architecture as there are plenty of n^2 operations on the length of the input [2]. Luckily for us, a bunch of companies started releasing for free giant models trained on lots of data, researchers learned how to "fine tune" them to specific tasks using way less data than what it would have taken to train from scratch, and transformers exploded in popularity.

[1] I use "mean" in quotes here because the transformer can only learn from word co-occurrences. It knows that "grass" and "green" go well together, but it doesn't have the data to properly say why. The paper "Climbing towards NLU" is a nice read if you care about the topic, but be aware that some people disagree with this point of view.

[2] The transformer is less efficient that an LSTM in the total number of operations but, simultaneously, it is easier to parallelize. If you are Google this is the kind of problem you can easily solve by throwing a data center or two at the problem.

I'm the author of https://jalammar.github.io/illustrated-transformer/ and have spent years since introducing people to Transformers and thinking of how best to communicate those concepts. I've found that different people need different kinds of introductions, and the thread here includes some often cited resources including:

https://peterbloem.nl/blog/transformers

https://e2eml.school/transformers.html

I would also add Luis Serrano's article here: https://txt.cohere.com/what-are-transformer-models/ (HN discussion: https://news.ycombinator.com/item?id=35576918).

Looking back at The Illustrated Transformer, when I introduce people to the topic now, I find I can hide some complexity by omitting the encoder-decoder architecture and focusing only on one. Decoders are great because now a lot of people come to Transformers having heard of GPT models (which are decoder only). So for me, my canonical intro to Transformers now only touches on a decoder model. You can see this narrative here: https://www.youtube.com/watch?v=MQnJZuBGmSQ

- You can develop a very deep understanding of a sequence by observing how each element interacts with each other over many sequences.

- This understanding can be encapsulated in "compressed" low dimensional vector representation of a sequences.

- You can use this understanding for many different downstream tasks, especially predicting the next item in a sequence.

- This approach scales really well with lots of GPUs and data and is super applicable to generating text.

Transformers are about converting some input data (usually text) to numeric representations, then modifying those representations through several layers to generate a target representation.

In LLMs, this means go from prompt to answer. I'll cover inference only, not training.

I can't quite ELI5, but process is roughly:

  - Write a prompt
  - Convert each token in the prompt (roughly a word) into numbers.  So "the" might map to the number 45.
  - Get a vector representation of each word - go from 45 to [.1, -1, -2, ...]. These vector representations are how a transformer understands words.  
  - Combine vectors into a matrix, so the transformer can "see" the whole prompt at once.
  - Repeat the following several times (once for each layer):
  - Multiply the vectors by the other vectors.  This is attention - it's the magic of transformers, that enables combining information from multiple tokens together.  This generates a new matrix.
  - Feed the matrix into a linear regression.  Basically multiply each number in each vector by another number, then add them all together.  This will generate a new matrix, but with "projected" values.
  - Apply a nonlinear transformation like relu.  This helps model more complex functions (like text input -> output!)

At the end, you'll have a matrix. You then convert this back into numbers, then into text.