Transcript for:
Long Short Term Memory (LSTM) Overview

Imagine you're at a murder mystery dinner. Right at the start, the lord of the manor abruptly keels over, and your task is to figure out who done it. Could be the maid, could be the butler. But you've got a problem.

Your short-term memory isn't working so well. You can't remember any of the clues past the last 10 minutes. Well, in that sort of situation, your prediction is going to be, well, I think... better than just a random guess. Or imagine you have the opposite problem where you can remember every word of every conversation that you've ever had.

If somebody asked you to outline your partner's wedding vows, well, you might have some trouble doing that. There's just so many words that you'd need to process. Be much better then if you could just remember, well, the memorable stuff. And that's where something called Long short term memory comes into play, also abbreviated as LSTM. It allows a neural network to remember the stuff that it needs to keep hold of context, but also to forget the stuff that, well, is no longer applicable.

So take, for example, this sequence of letters. we need to predict what the next letter in the sequence is going to be. Well, just by looking at the letters individually, it's not obvious what the next sequence is. Like we have two Ms and they both have a different letter following them.

So how do we predict the sequence? Well, if we have gone back through the time series to look at all of the letters in the sequence, we can establish context and we can clearly see, oh yes, it's my name is. And if we, instead of looking at letters, looked at words, we can establish that the whole sentence here says, my name is, oh yes, Martin.

Now a recurrent neural network is really where an LSTM lives. So effectively an LSTM is a type of recurrent neural network. Recurrent neural net.

And recurrent neural networks work in the sense that they have a node. So there's a node here and this node receives some input. So we've got some input coming in.

That input is then processed in some way, so there's some kind of computation and that results in an output. That's pretty standard stuff. But what makes an RNN node a little bit different is the fact that it is recurrent and that means that it loops around. So the output of a given step is provided alongside the input in the next step.

So step one has some input, it's processed and that results in some output. Then step two has some new input but it also receives the output of the prior step as well. That is what makes an RNN a little bit different and it allows it to remember previous steps in a sequence.

So when we're looking at a sentence like my name I, we don't have to go back too far through those steps to figure out what the context is. But RNN does suffer from what's known as the long-term dependency problem, which is to say that over time as more and more information piles up, then RNNs become less effective at learning new things. So while we didn't have to go too far back for my name I, If we were going back through an hour's worth of clues at our murder mystery dinner, well that's a lot more information that needs to be processed. So the LSTM provides a solution to this long-term dependency problem, and that is to add something called an internal state to the RNN node. Now when an RNN input comes in, it is receiving the state information as well.

So a step receives the output from the previous step, the input of the new step, and also some state information from the LSTM state. Now what is this state? Well, it's actually a cell. Let's take a look at what's in there.

So this is an LSTM cell, and it consists of three parts. Each part is a gate. There is a forget gate.

There's an input gate and there's an output gate. Now the forget gate says what sort of state information that's stored in this internal state here can be forgotten. It's no longer contextually relevant. The input gate says what new information should we add or update into this working storage state information.

And the input gate says what the current state information is. Output gate says of all the information that's stored in that state, which part of it should be output in this particular instance. And these gates can be assigned numbers between 0 and 1, where 0 means that the gate is effectively closed and nothing gets through, and 1 means the gate is wide open and everything gets through.

So we can say forget everything or just forget a little bit. We can say add everything to the input state or add just a little bit. and we can say output everything or just output a little bit or output nothing at all. So now when we're processing in our RNN cell, we have this additional state information that can provide us with some additional context.

So if we take an example of another sentence like Martin is buying apples, there's some information that we might want to store in this state. Martin is most likely to derive to the gender of males, so we might want to store that because that might be useful. Apples is a plural, so maybe we're going to store that it is a plural for later on.

Now as this sentence continues to develop, it now starts to talk about Jennifer. Jennifer is... At this point we can make some changes to our state data. So we've changed subjects from Martin to Jennifer, so we don't care about the gender of Martin anymore. So we can forget that part, and we can say the most likely gender for Jennifer is female and store that instead.

And really that is how we can apply this LSTM to any sort of series where we have a sequence prediction that's required and some long-term dependency data to go alongside of it. Now, Some typical use cases for using LSTM, machine translation is a good one. And another one are chatbots, so Q&A chatbots, where we might need to retrieve some information that was in a previous step in that chatbot and recall it later on. All good examples of where we have a time sequence of things and some long-term dependencies.

Had we also applied LSTM to our murder mystery dinner, we probably could have won first prize by having it forecast to us that whodunit was the butler. It's always the butler. If you have any questions, please drop us a line below.

And if you want to see more videos like this in the future, please consider liking and subscribing.