AWS AI & Machine Learning Podcast

Episode 15: AWS news

April 07, 2020 Julien Simon Season 1 Episode 15
AWS AI & Machine Learning Podcast
Episode 15: AWS news
Show Notes Transcript

In this episode, I go through our latest announcements on AWS DeepComposer, Amazon Transcribe Medical, and Amazon Personalize.

⭐️⭐️⭐️ Don't forget to subscribe to be notified of future episodes ⭐️⭐️⭐️

Additional resources mentioned in the podcast:

For more content:

  • AWS blog:
  • Medium blog: 
  • YouTube: 
  • Podcast: 
  • Twitter

speaker 0:   0:00
Hi, everybody. This is Julian. Formidable. Yes. Welcome to Episode 15 of my podcast. Don't forget to subscribe to be notified of future episodes in those difficult times, I hope you're doing okay as good as possible. And I hope you're safe in this episode. I'm going to focus on the latest news from edible US on machine learning. And we're going to look at high level service is mostly deep composer transcribe, medical and personalized. So let's get started. S so big news on Deep Composer this week. The service is now generally available, and we added a whole bunch of feature. So if you remember, the composer was launched in preview, actually invent 2019 and s, so you could use it in the Consul. You could use either the virtual keyboard or the real keyboard, this one, and you could record some tunes and use pre trained models to generate compositions. So now you can only use Deep composer again. You don't need to have a keyboard. We'll get back to that. But you can very well start with the virtual keyboard in the council. And we added some features. So, uh, we have some Ah, what we call the learning capsules. So basically a short lessons generated A I and generated adversarial networks introducing the theory what those things are, how to train them, how to evaluate them. And this should give you confidence to move on to in council training because now you can train your own models with composer. Okay, I'm going to show you the consul in a second. So we have some easy steps to train a model without writing a line of machine learning coaches working in the council. And then you can use your own models to generate compositions. Okay. And, uh, so the physical keyboard is now available for sale on amazon dot com. And if you buy one and if you link the keyboard to your universe account, you get three extra a month off free trial, so I guess, as an incentive to buy the keyboard. But again, you don't need to buy the keyboard. You can use the virtual keyboard in the council, or you can use pretty much any media keyboard out there, and it should work fine. But hey, you want extra, uh, extra time to train your models, then buying the keyboard. It might be an interesting option. Okay, so if we move to the deep composer counsel, we can see a few more things. Okay, Uh, so if you've never seen that thing before, it is okay, so you can play music. We have some pre recorded tunes that you can generate compositions for. Or of course, you can click on the virtual keyboard. All play on a physical keyboard and record tunes. All right. But I guess the new thing, the main thing is creating a model. So you would just go here, create a model, select one of those two Ganz architectures, some details. And don't worry. I mean, if you're not familiar with this, this is exactly what the learning capsules are for. They will explain what those things are and how they work. Then you select the data sets to train the model on your symphony and jazz pop rock. You can set some hyper parameters, but you can leave those as is, I guess, for the first few trainings. Don't go wild and just use the defaults. Give the model name and click, and off it goes and it's going to train for about eight hours and then you can use that model Thio to generate composition. So this is a really similar experience to ah, Deep Racer. If you use depressor before to train reinforcement learning models for the tournament driving car, this is really similar. Just working the council tweak parameters of it and try to generate interesting models. All right, so again, this is not available to everyone. And, uh, and you can have fun playing with music. And no, I'm not gonna play anything because I literally cannot play the keyboard or anything else for that matter. So that would be an absolute disaster. But go and check out deep composer and place a music. All right. Okay. The second announcement is on. One of my favorite service is and this service is transcribed medical. So transcribe medical was also launched training in 2019 and it's basically an extension of transcribed, transcribe, bizarre speech to text service and transcribe medical is an extension of that for medical vocabulary. And well, I did cover this service quite a lot, and I will have links to the block post and to this short video demo which are recorded at the time showing me reading medical text, which I have no idea what it is, really. But the important bit is transcribed. Medical picks it up perfectly, and it's quite impressive. So the announcement here is that you can now do batch transcription. So if you have a bunch of audio files with either medical conversations or medical dictation, you can just upload them to us three and launch a transcription job. Okay, so this is really as easy as this. Okay, input your data in unnecessary bucket. Aah! Select either conversation type or dictation type. And I'm guessing the only difference here is going to be the number of speakers. So if you have conversation that transcribed, will that work on dry to identify different speakers versus dictation, which should only be one person speaking. OK, so input data on his three and then out. Good data in history, OK, and you can encrypt results with Kms if you want to keep all those medical documents safe. And of course, you should write so critical new feature from transcribe medical just making it easier and easier to process magical documents. That skill and the 3rd 1 I want to talk about is another off My third service is a meson personalized. So personalizes a high level service that lets you build easily are personalization and recommendation models. And again, I've covered this quite a bit in the past. And now we're happy to launch recommendation scores in personalized. And there's a really nice block post from my colleague Brandon. Here, I'll include a link to that. So let me let me explain what this is so previously when you trained a recommendation model, Okay, the high level process would be upload a data set to his three. Okay. From the minimal data said was the user item interaction data said pretty much C S V file showing that user 123 as ah interacted with the item 456 etcetera, etcetera. OK, so showing interaction between users and items which could be anything movies you like, songs you like product you bought, you know, anything you want. And so then we would train Ah model which personalized calls a solution, okay. And the solution would be based on a recipe and a recipe is more than on algo. A recipe is off course on algorithm for the specific problem you're trying to solve, But it's also pre processing steps for the data. And it's, ah, tuning steps to optimize the accuracy and the machine learning performance of your model. Etcetera, etcetera. OK, and you could either pick a recipe yourself or you could use photo email and, um, and just let personalized figure out which recipe would work best. Okay, so here's the Here's when I tried a while ago and here I used this h r n n I'll go, which is a deep running. I'll go for recommendation. And as you can imagine from the name, it's based on the recurrent nor let words. Okay, so I trained it on this data set, and then I deployed it to campaign. Um, and we could run some predictions there. Okay, So using that campaign, and I couldn't have asked for prediction and just under a user, I d here and and get some recommendations. And I would get a list off in this case movies because this is a movie. Is there to say I would get a list off movies that this user might like. Okay, but I wouldn't get any other information besides Thea, Uh, the movie I d thanks to Disney feature. Now we have recommendation scores and recommendations scores. We'll let you know which are the top items that user may enjoy. Okay. And as you saw, these are pretty tiny scores. Because, in fact, um, each item in the data set has its own score and all this course add up to one. Okay, so if you have thousands and thousands of items and the day said you will end up having tiny scores, but the scores are related to one another. Okay, so if you use personalization, each item gets a score and the old Adam to one. If you're working with the ranking model than the score will be higher, because we will only generate scores for the items that need to be ranked. So it's gonna be a subset of the data set, and obviously, we will have Ah ah, fewer items there and hence the score will be will be higher. Okay. And machine learning. Aah! Freaks will recognize the self max function, which we use a lot in in she learning and deep Ronnie. Okay, so each score is between your anyone and they all add up to one, okay? And The cool thing is, aside said, this is really ah, model that I trained ages ago, and I didn't have to retrain it. So I just run that demo again and voila, right? And we get scores. So if you have models that you already deployed can already campaigns that are already running, then chances are you don't need to do anything right. A cz long as you use one of the al goes that actually supports those. Ah, those scores right, which is mostly the hr nn and a few more Then you should be fine. Just, ah, run predictions and you will cease course popping up. Okay, so this is really cool. This is Ah, one of the top requests that I got on personalized. So should make a lot of people happy. Hopefully. All right, well, that's it for personalized. All right, that's it for this episode again. Please subscribe to my channel and please stay safe. I hope to see you on the road pretty soon. Until then, keep product