So first of all, before I start, I really want to think everybody involved in making just confusion happen for having me and for really organizing a really cool event so far.

And thank all of you for being here this morning.

It might be a bit early for some of you.

So thanks for coming here to you.

Listen to me.

Talk about machine learning forefront and developers.

So before I stop, I've given this talk a couple of times before, and the feedback or comment usually get is that I speak very fast.

So I'm gonna try to be mindful and slow down a little bit on some slides.

But I get really excited about sharing what I know with you saw.

My natural pace always comes back on.

And also, I mean, uh, I'd like to remind you that we're in a fast paced industry, so I'm not, You know, I'm not fast.

You're slow.

So I'm getting everything on a try.

But so my name is Charlie Gerard.

I am a developer at a product company called at last year in in Sydney, and outside of that in the community.

I'm also part of to take community programs called Google developer Experts and magilla take speakers.

If you're interested in learning more about what it is, I know that there's a few others here as well.

So there's quite a few community of people sharing knowledge through conferences and look posts and workshops and things like that.

But I'm not here to talk about that.

I'm here to talk about machine learning, so usually I start because I don't know what you know and what you don't know.

So I usually start with an example to try to put everybody on the same level about what machine learning is.

But yesterday I think Money Car did a very nice introduction, talking about how to recognize our pictures of cats and dogs, and I thought that was really simple and well done.

So I kind of remove some of my slides.

And I'm just gonna briefly say, if you were not here or if you, you know, don't remember that it's basically giving the ability to computers to find patterns in data without being explicitly programmed.

So if I just take back the example from Monica yesterday, the first time that a computer sees a picture off a cat.

It doesn't really know what it is.

It's never seen one before.

But you know, by giving it millions of samples of data off what a cat looks like in a dog looks like using algorithms and machine learning models, the computer is able to understand, to find patterns in what makes a cat a cat on a picture and what makes a dog a dog.

So at a high level, this is basically that.

So at the core of it is machine learning models, and to build a model, you need algorithms.

So let's dive a little bit into this.

So if you start looking into machine learning, you might come across some of these algorithms.

So I know that I when I started, I definitely heard a lot about my face and Kenya Rhys neighbor.

And there's really a lot of them, and they're good at solving different types of problems and dealing with different types of data.

So before even looking into what algorithm you want to use, you have to understand what type of problem that you want to solve.

So there's three main types of problems.

Problems in machine learning.

You have supervised learning, unsupervised learning and reinforcement.

Learning for this talk, knowing that only have half an hour.

I'm only gonna cover, supervised and unsupervised.

Also, because a lot of the time when you get started, the problems that you want to try and solve can be solved with supervised and unsupervised.

So at first, maybe don't bother with understanding what reinforcement is.

So let's start with supervised learning.

If we start with a really short definition, it's basically creating a predictive model based on a set of features and labels.

And there's three main terms in this definition that you need to understand.

It's, you know, what is a predictive model?

What's features and what's a label.

So I don't know about you, but I learn more when I can kind of, like relate to examples.

So a classic example of a supervised learning problem is predicting the price of the house.

So let's imagine, you know, we're in Singapore.

We have a house if we're lucky enough that we want to be able to sell, and we don't know how much we should sell it for free to, you know, for us to make perfect and you know said it fast and things like that.

But what we have is a lot of data of all the houses in Singapore, and that's what we have.

So we start with that.

So how can we predict the price off our house based on the price of houses in the market?

So the label is in that case is the price.

It's what you're actually trying to predict.

So in your data set that you're going to put together, it's Hollywood Classify each entry.

So as our problem is trying to predict the price we would have.

If, you know, imagine an Excel spreadsheet.

We would have a condom with all the prices off all the houses in Singapore, and then we would have.

Our features and features are the characteristics of the entries in your data set.

So four houses it would be the number of floors, the number of rooms.

Bathrooms doesn't have a garden.

Oh, no, the neighborhood and things like that.

And using these labels and features or using an algorithm, you would try to understand the correlation between the features and the price for the characteristic and the price of the house to generate a model that can then end up being a mathematical representation on the outcome of your training.

So using everything that you know about the houses use algorithms that end up being like a function that takes new parameters of a new house.

It's never seen before, and it's able to generate ah prediction off a price that the house should go for, based on all the houses that it's seen on the market.

So it's important to know that supervised learning needs labels and features to generate a predictive model.

And then, if we look at a type up, another type of learning the unsupervised learning, it's the ability to create predictions based only on a set of features.

And the end of the definition is really important because it's the main difference between unsupervised and supervised.

So if we think about our problem of predicting the price of a house, it wouldn't really work as an it's not an unsupervised problem.

How can you understand?

How can you predict the price of a house if you have never even seen what the price is?

You only know the characteristics of a house.

So in that particular case, a type of problem that unsupervised learning solves is more around clustering, so an example of that would be predicting customer behavior.

So when you go to the supermarket, they might not know who you are at all.

But based on what you buy, how often how expensive it is, where they can actually really cluster people together into different categories and based on your buying behavior, they would then be able to know what advertising to send to you because you're more likely to buy the same thing that people like you have bought as well.

So so unsupervised learning is more around clustering rather than getting a particular answer at over question.

So in this, too.

So now that we talked about the types of learning I talked about, you know, problems that you're predicting the price of a house or customer behavior.

But it might not be something that you want to do.

And yesterday in Monica stock, we saw that machine can be used in a lot of more creative ways.

So let's talk briefly about different applications.

I just put like a short lease together, and the 1st 1 I'm going to talk about is this example here where you upload a picture of a cat and then on algorithm generates for you a piece of text that you can use as all text for your image tags in your is Jamil.

So this experiment was built by Sarah Dresner.

I put the link there if you want to try a bit later.

But so maybe for one image, my deal.

Oh, I can do it myself.

But when you like, you can see that using machine learning can maybe eventually automates some of the things that we do as developers.

So we can actually spend time, um, solving some of the logic that we want to solve.

So that's just like a quick example.

Another one that would have been really useful for me a few years ago I was.

It's this framework called not safe for work Gs where again you upload a picture in in this little experiment here, and it gives you the probability of being an image that is not safe for work.

And I used to work in advertising, and when you work for brands, sometimes they want to put competitions together where they want to engage with customers.

So people have to upload a selfie or whatever to win a contest.

So people approved a lot of images, and most of the time they brain wants to see the images live in a gallery or whatever.

And the thing is that the check that you have to do to make sure that these images are okay they usually don't manually, either by by like a product manager or ah or by the developer is so using machine learning in the front end.

You could actually do a pre check of what people are applauding before saving into the database and before displaying it live on a page.

So that's definitely something I could have used a few years ago.

So I thought it was really interesting.

And then fight like another example that is really, really, like, totally like, awesome, like I really love it because it makes is a lot of different technologies, and it can actually help people in terms of accessibility.

So I don't know if you've seen this one, but this developer called Baby Shake Sing created a prototype around the concept of kind of like future interactions.

So he was thinking about if no voice you eye when you talk to Amazon.

Alexa, what about people who can't talk?

Are they?

Are we just gonna totally cut them out of, you know, innovation and things like that?

So what he decided to do is to use machine learning in javascript to look at the to capture data from the web cam feed, train an algorithm to recognize and gestures map them two words so he could actually communicate with Alexa just via via gestures like by training and algorithm.

So I thought it was really a creative way and really interesting way of using MACHIN owning and jealous.

So let's dive a little bit into instead of just talking about what you can do.

Um, let's have a look a little bit about.

First of all, why?

Because when you think about machine learning, you probably think about Python before you think about JavaScript.

But I think it's really interesting to now be able to do that in JavaScript because in terms of learning curve, it's a lot easier for developers to know that they only have to understand a few concepts of machine learning and, um, and understand the syntax off in framework rather than having to learn a new, entirely new language.

We learn new syntax, like every day when we move from Jake, where you too angular to view to react so we could have used to doing that.

And now it's just transferring that skill of being, you know, adaptable to another domain.

So I really love the fact that you don't have to be that a scientist.

You don't have to only know Python.

It's actually out there and you can start doing it now.

You know, the point that I think is important is around rapid prototyping.

So I really love doing that.

It's, you know, like this.

What I'm talking about has nothing to do with what I do for work.

So all of my experiments are things that I do outside, you know, after work, you know?

Yeah, or on the weekend.

So sometimes when I want to valued an idea I needed to be fast because I don't want to spend I'm not that patient.

So I don't want to spend six months on the project.

Usually want to be in, like, two hours.

Eso being able Thio two prototype something fast in Java script is really cool, because if you want to, just, you know, if you have an idea in mind, and often if you want to be given the time or the money at work to exploit more, you have to convince people that it's worth it and to be able to convince them.

Usually you have to show them something some kind of M V P or a proof of concept, and ah, and you being able to do machine learning And JavaScript means that you could valid an idea and then eventually moved to Python.

If you want to push it further.

But I think that's one of the really, like, awesome thing that I like about you.

And then finally there is a bigger and bigger ecosystem of tools that you can use in jail a script, and there's more and more documentation.

There's a lot of tutorials and courses, so I think it's a good time to get started if you want to get into that space.

So what can you do so you can do three things?

First of all, you can import on existing pre trained model, so it's a model that has already been trained with a lot of data for a specific purpose, and you can just imported in your app and use it as is.

You can also do.

You can also retrain and unimportant model that we call transfer learning, so you kind of do the same thing from the step one you can you can use.

The model of that is open source.

You can as you you can add your own samples on top of it and kind of retrained that really fast.

And then, finally, you can do all of it in JavaScript in the browser you can defy in train and run your model entirely in the browser.

So in terms of tools, I think the most popular is tensorflow Js.

But sometimes I feel like some of the terms that tensorflow Jess uses might be confusing for people who is just getting started.

So I really like the 2nd 1 on the right ml five.

Jess that it's a framework that tries to abstract some of the terms like, you know, it's trust to make it really, really easy for you to get started.

So it's a lot simpler, and as a few examples I feel like it's No, it's still kind of like in construction.

There's some examples on something a great things that work.

But sometimes when you look at the docks, you can see that it's still in progress.

Um, I definitely have a look.

And then if you don't really want to write an equally but just wanted paying some AP eyes, you can leverage the tools from Amazon, Google and Microsoft, and then the the M in the middle is for much energy.

Es eso.

Now, if you were here yesterday, you know what that is?

But yes.

So it's like some some frameworks are actually more specialized about certain type of data and to do certain things already.

Okay, so we are at this part where I'm going to do some demos and code because I feel like to get people on board.