Great post and write-up - I also made an in-depth explorations and did my best to use visuals - for anyone interested you can find it here: https://photonlines.substack.com/p/intuitive-and-visual-guid...
The Illustrated Transformer (2018)
11–14 of 14 posts
Re: The Illustrated Transformer (2018)
#12I gave a talk on using Google BERT for financial services problems at a machine learning conference in early 2019. During my preparation, this was the only resource on transformers I could find that was even remotely understandable to me. I had a lot of trouble understand what was going on from just the original publication[0]. [0] https://arxiv.org/abs/1706.03762
Maybe it's easier to understand in the format of annotated code https://nlp.seas.harvard.edu/2018/04/03/attention.html
Updated (in case above link goes away) - https://nlp.seas.harvard.edu/annotated-transformer/
Thanks for original!
Re: The Illustrated Transformer (2018)
#13This is good, but what bade me finally understand the transformer architecture [0] and attention [1], are 3Blue1Brown's videos.
Re: The Illustrated Transformer (2018)
#14While I absolutely love this illustration (and frankly everything Jay Alammar does), it is worth recognizing there is a distinction between visualizing how a transformer (or any model really works) and what the transformer is doing. My favorite article on the latter is Cosma Shalizi's excellent post showing that all "attention" is really doing is kernel smoothing [0]. Personally having this 'click' was a bigger insig…
Interesting read! As a mathematician I always had difficulties with AI jargon, even though I've been writing on neural networks since the late nineties.
Almost at the same time as the emergence of transformers, I had only minimal contact with the field. I was just aware of the appearance of the Vaswani paper, but only now have I returned to the subject in a way that requires more rigour. And I stumbled upon "attention" in the same way as the author. It did not help to know more about the biological model [1].
Yes, kernels. Asking myself, what software implementations of the superior colliculus or yet retina cell complexes like DS (direction-sensitive) or OMS (object-motion-sensitive) cells could provide.
[1] for example: https://mitpress.mit.edu/9780262019163/the-new-visual-neuros...