Home Transformers are LSTMs v2

Transformers are LSTMs v2

So I recently got into a bit of an argument with 2 friends while driving in the car to get dinner. It went a little bit like this:

Me: I think LSTMs are definitely the precursor to transformers, minus all the parallelization.

Henry: What?!

Me: Attention definitely came from transformers.

Raghav: I think if you told anyone who hasn’t thought insanely deeply and seen the weird connection, that LSTMs are transformers, they’d think you’re an idiot.

Henry: You can definitely do LSTMs with attention.

The rest was irrelevant banter, but TLDR: I still believe LSTMs are an early form of attention. I’ll just drop this graphic that I made here and let you draw the connections. I also wrote a medium article on an LSTM deep-dive years ago so I’m pretty convinced that I know a thing or two about LSTMs.

LSTMs and Attention

I’m fully aware of all the different types of attention exist (here’s a really nice post by one of my favorite technical writers, Lilian Weng, that summarizes them) and LSTMs don’t fit squarely into any of those boxes, BUT, the purpose and general mechanism of all these gates still very much are those of attention mechanisms.

This post is licensed under CC BY 4.0 by the author.
Recently Updated