How does the ELMo machine learning model work?

June 12, 2019 / Ask Slater, Machine Learning

At its core, ELMo is an RNN model with an LSTM gating setup that’s trained as a bidirectional language model.

But that’s not really what you asked. You asked how it works. It’s often really tough to explain how a machine learning model works. Importantly you have to understand that no part of what I said above is unique to ELMo. To understand why it “works” you have to understand what they changed from previous approaches, get pretty far into the weeds on exactly how they implemented this, and look at some benchmarks that show you exactly how well it works in which contexts.

Man, that sounds like a lot of work though. Don’t worry, someone already did it for you: Deep contextualized word representations

You see, the nice thing about Machine Learning is that it’s arguably the most open field of study that has ever existed. This is partially because it’s driven in large part by industry labs that have a large vested interest in keeping this research open. This field is driven by arxiv and conferences, not elsevier. That means that basically any advance in the field that’s important is public (including ELMo)

Not only have they made public all of their research, but they’ve actually released code and pretrained models along with it. Everything you can ask for. Unfortunately machine learning isn’t a trivial space to learn. If the paper I linked above doesn’t make sense then start with my first sentence. Break it down and learn the piece parts (Recurrent Neural Network, LSTM, gating, language modeling, bidirectional language modeling) even that though requires a pretty significant math background.

If you don’t understand the paper and you don’t have the needed math to understand the terms I’ve laid out above you’ve got two options:

  1. Learn the math. Once you’ve gotten through linear algebra, partial differential equations, and basic graduate-level statistics then this should be approachable
  2. Tell everyone that deep learning is an evil black box that should never be used and that it’s impossible to understand.

Most people in this spot pick #2. Hopefully you pick #1.

View original question on Quora >

Follow Slater on Quora >>

Don't Miss a Post!

Subscribe to indico's monthly newsletter to receive the latest blog posts and AI industry news.