Technology 4 Learning logo

Technology 4 Learning

Technology 4 Learning

Telephone1300 323 232

Emailt4linnovations@det.nsw.edu.au

Episode seventeen transcript

The Virtual Staffroom Podcast

Episode 17 – Cooking with Artificial Intelligence

 

Joachim Cohen:

Welcome to the Virtual Staffroom, a podcast made for teachers by teachers and all with a dash of educational technology thrown in. My name is Joachim Cohen, and today I'm joined by one other awesome member of the Technology 4 Learning Team, Linda Lazenby. Welcome Linda.

 

Linda Lazenby:

Hi Joe.

 

Joachim Cohen:

Now, you might've been wondering where we've been. It's been a little while since our last podcast. Well excitingly enough, we have been hitting the road, heading west and having a blast with the students and teachers in Regional New South Wales public schools, but we're back. And so what do we have in store for you today? Well, we're heading to the kitchen, but not your average one. Today, we are going to see and learn about the power of AI with someone who has led its potential loose on recipes. We are going to cook up a storm of awesome.

 

Joachim Cohen:

What do you get when you cross Marmite with Maltesers? Loss for words? Well, I think we can all understand. But our guest today has used the power of AI to see what type of awesome recipe might result. Working deep inside the Google machine, Sara Robinson, artificial intelligence and machine learning specialist has been experimenting with AI in unusual ways, and ways which makes it easy for all of us to begin to understand what it is, what it does, and what its potential could be. Sara, my taste buds have stopped watering, but our brains are bursting with excitement. Welcome to the Virtual Staffroom.

 

Sara Robinson:

Thanks so much for having me. It's great to be here.

 

Joachim Cohen:

Sara, tell us about your journey and how you've ended up as an artificial intelligence and machine learning specialist at Google.

 

Sara Robinson:

Yeah, that is an interesting journey. I'll try to give the abridged version. So I'm a self-taught engineer. I didn't study engineering or computer science in school. When I graduated college, I studied business and international studies. I graduated college and I moved out to San Francisco having never been there before, kind of a bold move. And I was working at a startup incubator, helping them a bit with marketing and doing some research for them. And just being in San Francisco, there was a tonne of software engineers and developers and I did, before I moved there, I didn't know it was such a tech hub. Obviously sounds weird to say that now, but every other person I was meeting was a developer just everywhere. And so I started to get really interested in it and felt like they were speaking a different language from me. And so I decided I really wanted to learn to code.

 

Sara Robinson:

And so I just started teaching myself, and I was really lucky in the job I had at the time they were really supportive about giving me time to learn because these skills would be useful to my job. So I continued learning on my own and I eventually joined a startup called Firebase initially as their community manager. It's a database company. You may have heard of it, a lot more people have heard of it now than when I joined. And so yeah, Firebase eventually joined Google in 2014, so long time ago. So basically like as I was on the Firebase team, it's a great opportunity to just keep improving my programming skills because all of their users are developers. So to engage with the community, I had to be pretty well versed in the API, started just giving presentations on it. And when we joined Google, I stayed in that role for a little bit and then eventually moved to the Cloud team where I kind of fell into machine learning by accident.

 

Sara Robinson:

So repeated that process a bit, teaching myself machine learning on the job, just trying out products, reading tutorials, taking advantage of some internal resources at Google to, since there's so many talented ML engineers and data scientists here. And so yeah, just did a lot of learning on my own, through trying out our products and basically like a lot of my job involves teaching external people how to use our tools. And so in order to do that, you have to have a really good understanding of how they work. So that's kind of the short version of how I got to where I am. I can talk a little bit about what my role is today if that's helpful.

 

Joachim Cohen:

Oh yeah. That would be amazing actually, because I think the one thing that has struck me about the story that you've told just then is that oftentimes when people have to learn themselves and start from scratch, they're the best people to teach. And I really love your story about how it is that you've just jumped on any opportunity. Yeah. So tell us about your role now. What does it encompass?

 

Sara Robinson:

Yeah. You bring up a really good point. I feel like the reason that I really like creating content like blog posts and videos and tutorials, because I still remember what it was like to be new to a topic, and I try to bring that perspective to anything I create. So my role now, my official title is developer advocate. And in that role, I'm essentially an external facing engineer, helping our customers and anyone in the developer community that's using our ML tools. So it involves two things, external advocacy, which is anything you see that I create. So videos, blog posts, projects like the Maltesers one that you saw, tutorial sample code, giving presentations at events, things like that. And then the other part of the job is internal advocacy, which involves taking feedback from external users and bringing out to our product teams. We also sometimes call ourselves customer zero, which means that we're the first to try out some really newer products and provide feedback to product teams and use them from a customer perspective.

 

Joachim Cohen:

Oh wow. That is a really exciting sounding role and one where you can absolutely influence change in a really powerful way. And I think that really brings up something because you're educating people all the time. And I think that a lot of our listeners, they might be hearing those terms of machine learning and artificial intelligence and they're still a bit mystical to them. Do you have a really great way of kind of explaining what they are for our listeners and breaking it down?

 

Sara Robinson:

Yeah, definitely. So the way I like to think of it is you can think of it as like concentric circles, where the outer circle is AI, which stands for artificial intelligence. And this is essentially the practice of just enabling machines to mimic human behavior. And that can be done in many different ways. Machine learning is a subset of that, which essentially involves teaching machines to identify patterns by providing them data and allowing the machines to learn from that data to generate new predictions. The big difference there is that machines are learning from patterns in data that you provide rather than requiring you to write explicit rules or logic to program how to do something. So instead of writing a lot of different IF statements to distinguish something, let's say you're trying to write a program that can tell whether a picture is an apple or an orange.

 

Sara Robinson:

You could see how writing all the IF statements in the code to do that would get pretty unwieldy. If, you could look at the color, but then you might have black and white images, so you have to account for that. Different type varieties of both of those fruits. So that's where machine learning becomes really powerful because instead of explicitly programming rules, you are teaching a machine to learn from data.

 

Joachim Cohen:

Oh wow. That's a really great way of explaining it, I think. And I wonder, because we started this out with that amazing example you gave of Maltesers and Marmite, and I'm wondering, can you use those to tell us a little bit about, is that an example of a machine learning engine that you created, putting those ingredients in?

 

Sara Robinson:

Sort of. I can explain a little bit more about that project if you'd like, that sounds good?

 

Joachim Cohen:

Yes. Yes. Absolutely.

 

Sara Robinson:

Yeah. So the project originated because over the past year in lockdown, I got really into baking. I had always liked baking, but I was traveling a lot and didn't really have time for it. And so with the only place to go being my kitchen, I just started baking. I hopped on the banana bread bandwagon with everybody else, but I kept going, just got really into baking all sorts of things. And I quickly started to notice that baking recipes follow distinct patterns. And what I mean by that is there's like specific ratios of flour, fat, sugar, leavening agents that will distinguish like a cake from a cookie or a biscuit or bread. And so I wondered like, could I combine my day job and machine learning with baking somehow, because machine learning is all about identifying patterns and data.

 

Sara Robinson:

So I wondered, could you learn from these ratios? And if you look it up online, you can find cool diagrams that break down the ratios of these different types of baked goods. The idea is like if you memorise these ratios and many professional chefs know these ratios and they'll use that to improvise and build their own recipes. So I wondered if I could build a model to classify different recipes based on the amount of different baked goods. And I kept my model to just core base ingredients, things like different types of sugars, flour, butter, different leavening agents, salt, because these are really like what distinguished the recipe. So I didn't include inputs to the model like chocolate chips or sprinkles, because that actually might confuse the model and make it memorise something like, oh, if it has chocolate chips, it's definitely a cookie.

 

Sara Robinson:

So the model is just the base ingredients. And then once I had that, I kind of used that to... So the first model I built predicted whether a recipe was for a bread, a cake or a cookie, and Mars saw this project. I did a video with my teammate about it. I can send you a link if you want to include it in the show notes. And they were interested if we could use machine learning to make a recipe that uses one of their candy brands called Maltesers mostly sold in the UK. They are really delicious. I've tried some.

 

Joachim Cohen:

Oh, they are. We love them over here too. So don't worry-

 

Sara Robinson:

Oh you have them over there too, that's awesome.

 

Joachim Cohen:

Oh absolutely.

 

Sara Robinson:

Yeah. Anyone listening who hasn't tried them, definitely recommend it. It's a malted milk ball covered in chocolate. So when they approached us, I wondered, could I take the original model that I built, which basically took ingredient amounts as input and then outputted a prediction of the type of recipe? I wondered if I could flip that so that, input a type of recipe output the ingredient amounts. Another thing I forgot to mention of that original model, a cool thing I was able to do is I played around with the model inputs and tweaked it to get the model to predict that something was, 50% chance it was cake, 50% chance it was a cookie. And then I made it, I called it a cakie and it was really delicious. It was, it tasted exactly like half cake half cookie.

 

Joachim Cohen:

Wow. That sounds cool. Gosh, we need to see that recipe. Yeah. You're onto something.

 

Sara Robinson:

I'll send you the recipe. Yeah. So the Mars project. So I basically built an entirely new model and wanted to train it on British recipes because they use slightly different ingredients, probably more similar to what you use. Like self-raising flour is not something that's used at all here, but they use that in most of the recipes. Same with castor sugar is not used here at all, even though it's very similar to granulated sugar, we don't use that. So I wanted to make sure it was trained since it was primarily for a UK audience. I wanted to train it on British recipes. So we trained a new model on four different categories, cakes, cookies / biscuits, scones, and tray bakes. And then the model was able to give you the ingredient amounts for some, I think it was like 10 to 15 ingredients that are common among a lot of different recipes.

 

Sara Robinson:

So we use that to develop the base. And then one of the cool things about this project is it gives me the baker, a lot of creativity. So you have this base recipe, but you don't know like, should you make it in a cake pan? Should you make it as cookies? What temperature should you bake it at? Those were all human elements in the process. And so that's also the part where we decided how we would incorporate Maltesers into it. So we tried a bunch of different combos to end on something. And then the Marmite was a bit of a different story. Do you want me to go into that too?

 

Joachim Cohen:

Yeah. Yeah, I'm intrigued. It's like, how could you combine the two? Wow.

 

Sara Robinson:

Yeah. So we wanted to have... So what I ended up making was something that incorporated both a cake and a cookie recipe. They weren't together, it was like a cookie layer or biscuit depending on where you're from. Biscuit layer on the bottom and then a cake on top of that, with some Maltesers in the middle. And we wanted... The team that was working on this, we wanted to have one other fun twist. So being at Google, we found some search trend data and we noticed that a popular search in the UK was, is Marmite sweet or savory? And we thought it was really interesting that people weren't sure. And so we thought, could it be interesting to somehow incorporate this into the recipe? So the Marmite vision was not AI/ML related, and so what I did was I developed a frosting recipe to use as the topping for this whole thing that incorporated, it's kind of, it starts with a traditional buttercream base, but then it uses golden syrup, which is also really popular in the UK and delicious. Do you have that over there?

 

Joachim Cohen:

Oh yeah, absolutely. Indeed, it's my mouth water already.

 

Sara Robinson:

I love it. I order it all the time now. I don't know why we don't use that in the U.S.

 

Joachim Cohen:

On pancakes, it's delicious. Absolutely.

 

Sara Robinson:

Oh my gosh. I need to try that. Yeah, that sounds really good. So yeah, the frosting recipe, I think the only ingredients are patterned sugar, butter, golden syrup and Marmite, which sounds gross if like, depending on how you feel about Marmite, but it's actually delicious. I'm not even exaggerating. It tasted kind of like salted caramel.

 

Joachim Cohen:

This is exciting. This is exciting because we have Vegemite over here in Australia, you see. It's a bit like Marmite, and I love Vegemite and honey. So I think you've just sold all, at least sold me probably no other listener, but you've definitely sold me. This is exciting.

 

Sara Robinson:

That's awesome. I need to try Vegemite too. I looked it up before and found that it's not quite the same. So I'm going to order some and try it.

 

Joachim Cohen:

Oh, I can tell you, this is amazing because number one, we often use recipes as a way of thinking about computational thinking for teachers, and you have put it.

 

Sara Robinson:

Oh wow.

 

Joachim Cohen:

Yeah. In such an amazing way, especially the science behind it. It totally blows my mind. And I think teachers as well can go, "Oh, so AI is not going to come over and it's not designed to replace anything, it's designed to augment and help us out." And that's exactly what you've said. The baker is still the master, but there's this amazing AI element that can help us to get a good base to get started.

 

Sara Robinson:

Yeah. That's one thing I loved about this project is that I'm not professionally trained in recipe development at all. And so for me the hardest part is figuring out those core ratios that if I have an idea in my head of something you want to create, that's the hardest part for me and that's what the model is able to do. And then I was able to focus on the fun part of like interesting flavor combinations, like this fun Marmite thing, and figuring out what the end format would be and how to decorate. So yeah, they really augment each other in helpful ways.

 

Joachim Cohen:

Oh, unbelievable. And in your role you must get to see so many different forms of AI engines being developed and ML engines. What are some of the ways that our listeners could maybe visualize or see AI and ML working in the real world at the moment?

 

Sara Robinson:

Yeah, that's a great question. And AI is being used probably all around you in ways that you might not even recognise. So one is, if you have a photo sharing app on your phone and you're able to search that, it's most likely AI/ML powering that search. So let's say something and I search a lot, as cakes, or I can even get more specific of type of cake, layer cake, whatever. And it brings up photos of that or bringing up photos of your family and friends. So that's one everyday application of machine learning.

 

Sara Robinson:

Another is machine translation. So if you've traveled and used a translate app to either take a picture of a sign or a menu and translate that in real time, that's all powered by machine learning. Another example is in emails. So if you've noticed like Gmail suggestions as you're typing, that's another example and could keep going, voice assistance or chat bots, if you've ever been on a website and a little chat bot pops up, sometimes folks will use that, use machine learning in that to then direct you to the right person to help you with whatever you're looking to do. It's also used a lot in finance, healthcare, tonnes of different industries.

 

Joachim Cohen:

Wow. Wow. And that really does help me visualise it now that I know how it works and can see those applications and how I think the machine might be working a little bit. And I mean, there is a lot of fear out there about these kinds of tools taking away or replacing our jobs. But is it really a partnership, have you seen it and can see it being a partnership between AI and ML and the human factor?

 

Sara Robinson:

Yeah. I think it's definitely a partnership. One thing we talk about a lot is what's called like having a human in the loop, and you need this to, I showed that in my example of the baking, but I believe that this is really necessary in any application of ML. You need humans to verify that your model, whatever it's doing is working correctly, and models can also become non-performant over time. So like deploying a machine learning model is not just a one step thing, you don't just deploy it and you're done, because the environment around that model can change quite a bit. New terms can be introduced if you have a language model. And so you need to have people monitoring these models and humans are also incredibly involved in finding the data to train these models. And that's really an important step too, because you need to make sure that your data doesn't contain any biases and you take steps to minimise those biases if you've done all you can to collect data.

 

Sara Robinson:

So yeah, I would say the partnership between humans and machines, human in the loop is really essential to the process.

 

Joachim Cohen:

So interesting. Yeah, it is a machine, but really it is a person behind there that's making that machine work.

 

Sara Robinson:

Exactly.

 

Joachim Cohen:

Wow. Wow. Blowing my mind. That's for sure. And look, I know the teachers listening in to this will have lots of students out there that are going to be working alongside AI. What do you think that teachers and students can do now to help prepare for the future?

 

Sara Robinson:

Yeah, that's a great question. I think there's lots of great resources out there for students who are interested in learning to code. There's more and more resources we're seeing to help children learn those skills. And I can find some links and share with you in the show notes after when we publish this. So I think just encouraging students who are interested in that to start learning when they can. And similarly, there's lots of resources on machine learning as well. So similar to the way that I learned on my own, there's just so many great free resources online, and that applies to both teachers and students as well. So yeah, I can send you some specific examples to include in the notes.

 

Joachim Cohen:

Oh, that would be amazing. And I think, yes, again, your story is so inspirational on how there aren't barriers. You can go out there and just start learning and developing, following your passion and your interest. And I guess that's what I'd love to know next. What is next for Sara Robinson? What's on your plans? What are you doing next?

 

Sara Robinson:

So I am loving this like interaction between baking and AI and getting a lot of interesting responses on some of the content I've published around that. So, I would like to keep exploring that area. And as I mentioned, the project that I did with Maltesers, that was focused primarily on UK recipes. So I think you could train really interesting models based on different locations, because recipes vary a lot by culture. And also curious to know if this would work for cooking. I have a feeling it wouldn't work quite as well because baking is a very exact science, as I talked about with all the ratios, like if you change the amounts of butter or the amount of salt, your recipe will probably not turn out right at all. Whereas like in cooking, if you're making some... This isn't true for all cooking, but a lot of recipes are pretty forgiving.

 

Sara Robinson:

Like if you don't like the taste add a little bit more salt or if you don't like a certain vegetable, take it out and replace it with something else. So I think if it were to work for cooking, you'd have to do a very niche specific model and you'd need probably a lot more data to train it on since these ratios aren't as clear, but that is kind of an area that I'm interested in exploring for demos. Right now, we just launched a new ML platform on Google Cloud called Vertex AI. So I've been working on creating some training content for that. I'm excited to see if there's any potential for more baking demos to build on that. I'm always interested in new baking projects and how to combine that with machine learning. So if anyone has ideas, you can find me on Twitter, I'll drop that link in the notes. And yeah, we'd love to hear what other applications of ML people are interested in.

 

Joachim Cohen:

I think you'd get lots of people reaching out and I already want you to combine pavlova and Vegemite. So that's the thing, two big things in Australia, pavlova and Vegemite were my two thoughts. I wonder what might come of that.

 

Sara Robinson:

Yeah. I feel like my Marmite discovery basically, it's so salty that I basically just replaced it for salt. Usually buttercream is a heavily sweet frosting and I use it instead of salt and it kind of added this, it also added a really interesting texture. I don't know if Vegemite has the same texture that's kind of syrupy.

 

Joachim Cohen:

No, it's very salty though. It's probably even more salty than Marmite, but that's so interesting. I can see what you're talking about now, about it being such a formula. Wow.

 

Sara Robinson:

Yeah, because like sweet and salty are usually such a good combo, it's why a lot of people sprinkle sea salt on cookies. So yeah, I'd be really interested in trying it in pavlova. I haven't tried to make pavlova before, but maybe my first attempt will be with Vegemite.

 

Joachim Cohen:

We'll have to send you a bottle, that's for sure. Now Sara, we're getting to the end of our podcast today, but we have one question that we haven't prepared you for, that we give every single one of our podcast presenters. And that is something we call rocket ship robots. So you might have heard of a podcast called Desert Island Discs over in the UK. Where you have to choose the CD or the disk that you take with you on a desert island, but we're a technology podcast, so we're flying up in a robot into outer space. What piece of technology would Sara Robinson take with her?

 

Sara Robinson:

All right. I'm going to give kind of a boring answer, but probably my smartphone, just because it has a really good camera and I love taking pictures of things I bake, but also just anything. And then I could keep in contact with friends and family while I'm on this island. I like that question now.

 

Joachim Cohen:

Oh, and I like your answer and you're not alone with the smartphone, but we're always reminding people you've got to take the internet with you too. So you want to do all this stuff, you got to take the internet and your smartphone with you too. But Google has got an answer for that.

 

Sara Robinson:

You're right. There's no guarantee that there'd be a service on the island, but I'd at least have the camera still.

 

Joachim Cohen:

Exactly. Exactly. Oh, thank you so much, Sara. Thank you for some amazing out of the box thinking and helping us to get our heads around some of the awesome opportunities AI is going to present for our students and our teachers in the future, and good luck with your next recipe.

 

Sara Robinson:

Thank you. And thanks so much for having me on the podcast.

 

Joachim Cohen:

So Linda, what would you apply AI to streamline your classroom?

 

Linda Lazenby:

Look, I think it's really exciting to consider how we might be able to use some of our AI technology to support students in their learning in time rather than afterwards. So I'd be really keen to explore how we could do that a bit better. How about you?

 

Joachim Cohen:

Oh, look, I love that idea and I'd really like to continue it on and have an app that's connected to your amazing system that you developed. So that it gives the teacher a ping to let them know when someone's struggling. If they can see that they're maybe reversing, typing, reversing, typing, and the teacher can go out and give them a hand when they need, it's like a little red ping that pops up. So definitely working together.

 

Linda Lazenby:

There is so much potential, definitely.

 

Joachim Cohen:

Now thinking about what we have heard today, what are some resources students can explore tomorrow to explore the power of technology in aspects of their everyday life. Linda, did you have a resource to share?

 

Linda Lazenby:

I do. The Digital Technologies AI hub is a great resource to share with students when they're really trying to understand what is AI? How is it different to normal computing? There are some great resources on here around kind of understanding, but then also developing some of those skills, there's some unplugged lessons for teachers, and there's some other applications and games that you can use in your classroom to explore AI. How about you, what have you stumbled across, Joe?

 

Joachim Cohen:

Well, Linda, look, I love what you're saying because really AI is not something that most teachers would have learned about when they were in university. So it's a learning journey for them as well. And I love that Digital Technologies help for that reason, but I've also jumped in and found an amazing set of podcasts called the Edspresso Series, which have been created by another part of the New South Wales Department of Education, but open to everyone. And they interview all these experts in AI to think about, what is it? How does it apply in the classroom? How is it being applied in the real world? So I would totally go and take a listen to some of those to grow your knowledge so that you have a little bit of confidence talking about AI. And they're really exciting too. So if you've got a walk or you've got to commute, there's something to fill it.

 

Linda Lazenby:

Yeah. And if you're not multitasking, there are some great case studies as well on the hub that shows what schools are implementing too.

 

Joachim Cohen:

Oh, fantastic. Treasure troves. So whilst the avid listeners will know, we love to give you a voice. And so to close us out, here's a little jam of techno-wizardry wisdom. And today we have one we discovered in our recent trip to Dubbo.

 

Speaker 5:

From on the ground, hearing from some amazing teachers and being so inspired. So the top tip tip I've taken from all of the awesomeness that I've heard, and that is to share. So go on, share what you're doing at your next staff meeting, your next morning muster, create a buzz, show how awesome it can be, show how it can engage your students and how it can help them to achieve some amazing outcomes. And see how inspired it makes others. Over and out from on the ground, out here in Dubbo.

 

Joachim Cohen:

So Linda, did you learn something today?

 

Linda Lazenby:

Absolutely. Every single time.

 

Joachim Cohen:

Oh, my excitement, my intrigue, and my potentiometer in AI is growing and growing, I can tell you.

 

Linda Lazenby:

Just a little note, please be aware that all views expressed by the podcast presenters, that's us, are our personal opinions and not representative of the New South Wales Department of Education. Discussions aren't endorsements of third party products, services, or events. And please note that as much as we sound like it, we are not experts in legalese, tech speak or anything in between, we're just passionate people keen to boost technology for learning in the classroom and to help build the skills in your students and for you to solve the problems of tomorrow. Do your due diligence, read further, and if we've got something wrong, let us know. We too are always learning and always improving.

 

Joachim Cohen:

This podcast has been produced by the masterful Jacob Druce, with the assistance and supreme coordination of many more awesome members of the T4L team. Before we go, please make sure you send us through your comments, your word of techno-wizardry wisdom and your thoughts for new guests and segments. And if you liked the podcast, give us a rating, so more and more educators find us and be inspired to get a little techie in the classroom. Stay compassionate, stay curious, and don't forget to think computationally everyone. And thanks for joining us.