Why the original transformer figure is wrong, and some other tidbits about LLMs

https://magazine.sebastianraschka.com/p/why-the-original-transformer-figure

by rasbt • 3 years ago

237 49 3 years ago

magazine.sebastianraschka.com

amelius4• 3 years ago

Are there any still-human-readable pictures where the entire transformer is shown in expanded form?

dnw• 3 years ago

Not quite what you’re asking for but perhaps in the direction https://openai.com/research/language-models-can-explain-neur...

minihat• 3 years ago

Try this: https://jalammar.github.io/illustrated-transformer/

Attention is explained separately. I have not seen an all-in-one diagram and cannot imagine one being helpful, since there's too much going on.

chaxor3• 3 years ago

Do you mean a graph that contains all neurons from the network in one structure? Similar to these: https://gfycat.com/BonyTotalArthropods https://gfycat.com/BitesizedWeeBlacklemur

That would be wonderful and I have been trying to do this. However, unfortunately some 'assumptions'/shortcuts have to be made. For example, the attention matrix is not known without input, so if just the structure of the network (weighted by the weights) is wanted, you have to put in some value 'p' ('1', '-1', w/e) to these edges. Also skip connections have to be dealt with explicitly instead of just adding them to a block diagonal matrix as one would with an MLP.

I am very interested if someone has a good solution that already has done these things though.

amelius4• 3 years ago

Those are nice diagrams. Yes, well actually I'm interested in anything that's between the abstracted form in the paper and the fully expanded form where you see all neurons.

visarga6• 3 years ago

As someone who started reading ML papers 10 years ago, I find transformers pretty simple compared to most architectures. For example Google's Inception was a complex mesh of various sized convolutions, other models have an internal algorithm like Non Maximum Suppression for object detection, or Connectionist Temporal Classification for OCR, GANs use complicated probability theory for the loss function. Even LSTM is more complicated.

If anything, we have been abandoning exotic neural nets in favour of a single architecture and that one is pretty simple, just linear layers (vector-matrix product), key-value products (matrix-matrix product), softmax (fancy normalisation), weighted averaging (a sum of products) and skip connections (an addition). Maybe it's become hard for me to see what is complicated about it, I'd be curious to know what part is difficult. Is it the embeddings, masking, multiple heads, gradient descent, ...? Embeddings have been famous for 10 years, ever since the king - man + woman = queen paper. You don't need to be able to derive the gradients for the network by hand to understand it.

In short, a transformer is mixing information between tokens in a sequence and computing updates. The mixing part is the "self attention" or "cross attention". The updating part is the feed-forward sublayer. It has skip connections (adds the input to the output) in order to keep training stable.

chaxor3• 3 years ago

Even when someone understands the architecture very well, there's still great utility in having a full graph representation of the full NN architecture. This could be for teaching students, or for doing analyses on the structure of the full network, etc.

Buttons840• 3 years ago

Simple is good. Especially in machine learning where a bug usually means that it kinda works, but not as well as it could. Also, when an off-the-shelf algorithm half works, it's good to be able to add you own tweaks to it, and again, this requires simplicity.

For a complicated architecture to succeed, it's going to need to reliably achieve state of the art performance on everything without requiring any adjustment or tweaks.

rasbt3• 3 years ago

Agreed, compared to other architectures, transformers are actually quite straight-forward. The complicated part comes more from training it in distributed setups, making the data loading and tensor parallelism work due to the large size etc. Like the vanilla architecture is simple, but the practical implementation for large-scale training can be a bit complicated.

lalaithion• 3 years ago

The things that still confuse me about transformers is:

1. Why do we _add_ the positional embedding to the semantic embedding? It seems like it means certain semantic directions are irreversibly with certain positions.

2. I don't understand why the attention head (which I can implement and follow the math of) is described as "key query value lookup". Specifically, the Q and K matrices aren't structurally distinct – the projections into them will learn different weights, but one doesn't start out biased key-ward and the other query-ward.

kk582• 3 years ago

In CNN the layers seem to learn geometric primitive and deeper layers seem to learn more complex geometric patterns loosely speaking.

In transformer what do query key matrices learn? How are their weights somehow working to extract context no matter which word appears in which position?

qumpis• 3 years ago

King-man got me thinking about the VAE paper for a second :)

aptitude_moo• 3 years ago

I find this one as the most complete [1] [2]

[1] https://github.com/ajhalthor/Transformer-Neural-Network/blob...

[2] https://www.youtube.com/watch?v=Nw_PJdmydZY&list=PLTl9hO2Oob...

fn-mote• 3 years ago

This note contains four papers for "historical perspective"... which would usually mean "no longer directly relevant", although I'm not sure that's really what the author means.

You might be looking for the author's "Understanding Large Language Models" post [1] instead.

Misspelling "Attention is All Your Need" twice in one paragraph makes for a rough start to the linked post.

[1] https://magazine.sebastianraschka.com/p/understanding-large-...

visarga6• 3 years ago

which would usually mean "no longer directly relevant"

Or it could mean the lesson from these papers has been assimilated and spread wide and far, thus they are no longer "news". The pre-layernorm is one.

homarp3• 3 years ago

also one of these papers is from Schmidhuber

and https://news.ycombinator.com/item?id=23649542 gives some context to the "For instance, in 1991, which is about two-and-a-half decades before the original transformer paper above ("Attention Is All You Need")"

rasbt3• 3 years ago

Misspelling "Attention is All Your Need" twice in one paragraph makes for a rough start to the linked post.

100%! LOL. I was traveling and typing this on a mobile device. Must have been some weird autocorrect/autocomplete. Strange. And I didn't even notice. Thanks!

canjobear• 3 years ago

The original Transformer wasn’t user in an LLM.

andreyk• 3 years ago

The actual title "Why the Original Transformer Figure Is Wrong, and Some Other Interesting Historical Tidbits About LLMs" is way more representative of what this post is about...

As to the figure being wrong, it's kind of a nit-pick: "While the original transformer figure above (from Attention Is All Your Need, https://arxiv.org/abs/1706.03762) is a helpful summary of the original encoder-decoder architecture, there is a slight discrepancy in this figure.

For instance, it places the layer normalization between the residual blocks, which doesn't match the official (updated) code implementation accompanying the original transformer paper. The variant shown in the Attention Is All Your Need figure is known as Post-LN Transformer."

rasbt3• 3 years ago

So weird, I posted it with almost the original title (only slightly abbreviated to make it fit: "Why the Original Transformer Figure Is Wrong, and Some Interesting Tidbits About LLMs".

Not sure what happened there. Someone must have changed it! So weird! And I agree that the current title is a bit awkward and less representative.

trivialmath• 3 years ago

I wonder if for example a function is an example of a transformer. So the phrase "argument one is cat" and argument two is dog and operation is join so the result is the word catdog is operated by the transformer as the function concat(cat,dog). Here the query is the function and the keys are the argument for the function and the value is a function from word to words.

visarga6• 3 years ago

They can intelligently parse the unstructured input into a structured internal form, apply a transform, and then format the result back into unstructured text. Even the transform itself can be an argument.

ijidak• 3 years ago

Has anyone bought his book: "Machine Learning Q and AI"?

Is it a helpful read as a cliff notes for the latest in Generative AI?

YetAnotherNick• 3 years ago

This is the commit that changed it: https://github.com/tensorflow/tensor2tensor/commit/d5bdfcc85...

Why the original transformer figure is wrong, and some other tidbits about LLMs

Related Stories