AWS AI Machine Learning Podcast Episode 10 AWS news and demos

Transcript

Hi, this is Julien from Arcee. Welcome to episode 10 of my podcast. Don't forget to subscribe to my channel to be notified of future videos. In this episode, I will cover some latest announcements from AWS and share some interesting resources. Let's get going. Let's start with the regional expansion. In the last few weeks, we deployed AI services to additional regions. You can now use Amazon Comprehend in South Korea, Tokyo, and Mumbai. You can also use Amazon Forecast in South Korea, and Amazon Personalize in South Korea. So, it's a big week for South Korea users, right? And remember, these services will probably come to your regions as well. Keep asking, okay? The more people ask for them, the more chances they have of being deployed. Get in touch with your AWS contacts, or just keep posting in the AWS forums or on Twitter if you like. Now, let's look at some features that have been added recently. The first feature today is a new update on Amazon Personalize, our managed service to build recommendation and personalization models. You can now use up to 50 item attributes to build your recommendation models, up from five previously. This is a 10x improvement, meaning if you have wide datasets with various features, signals, and events showing user engagement with items from your collection, you can add more to the dataset and build more relevant models. That's pretty cool. We also added a really cool feature to Amazon Polly, our text-to-speech service. Customers can now get in touch with the Polly team to build a custom voice that represents their brand, company, or service. Let's listen to a couple of examples. The first one is the KFC voice: "Everybody loves Friday. Well, I'm not so sure about the joke, but I kind of like the chicken." That's one example. Let's listen to the other one. This one is for NAB, an Australian customer: "Welcome to National Australia Bank. This is the new voice of NAB, created by Amazon Polly, a service that turns text into lifelike speech. This voice has been uniquely created for NAB, providing a consistent experience for our customers whenever they call us." You can tell it's an Australian accent. Now, you can get in touch with us if you want to build a custom voice. It's one of our capabilities. So, that's really cool. Now, am I the one who's going to start the petition for a Jeff Barr voice? Well, I guess I am. If you agree with me, just keep tweeting. We need a Jeff. Sorry, Jeff. Let's talk about SageMaker. I'm not announcing any new features here, but I want to point you to a collection of new SageMaker notebooks put together by the SageMaker Debugger team. As you probably know, SageMaker Debugger allows you to inspect the internal state of your machine learning models, saving tensor information during training and then starting a debugging job or inspecting the tensors yourself. These notebooks do exactly that. If you go to the GitHub repository for SageMaker examples and zoom in on the SageMaker Debugger folder, you'll find these new examples. The tensor analysis and real-time analysis examples are absolutely amazing. There's a BERT example, a state-of-the-art natural language processing model, where you can visualize in real-time how the model is training. There's also a CNN model example that shows activation maps, highlighting which parts of an image are helping the model determine the class. This is trained on a traffic sign dataset, and you can visualize specific parts of the image that are helping the model figure out the class. This is super impressive work. I want to congratulate my colleagues who built these; they are really amazing resources. If you've never looked at SageMaker Debugger, please take a look at these examples. They really push the service to the limit, and you will learn a lot. The next item I want to talk about is not new, but it might have flown under everyone's radar. I'm talking about the NumPy API in Apache MXNet. Apache MXNet is an open-source deep learning framework, and AWS does a lot of work on MXNet. A few months ago, the project released a NumPy compatibility API, meaning you can use MXNet as a drop-in replacement for NumPy. Whatever NumPy code you have, you can just import MXNet.Numpy and use MXNet instead of NumPy. Why would you use MXNet instead of NumPy? Well, MXNet is much faster. It's natively written in C++ and makes good use of all your CPU cores or GPU cores if you have a GPU. In the launch blog post, there's a test where we multiply large matrices using NumPy, MXNet Numpy on CPU, and MXNet Numpy on GPU. MXNet NumPy on CPU is roughly three times faster than NumPy, and if you throw that matrix multiplication at an Nvidia GPU, it's blazingly fast. All it takes is importing MXNet Numpy and setting the GPU context in your NumPy calls if you want to use the GPU. These are really easy modifications, and if you're using NumPy at scale, this is definitely worth a shot. I'm not sure why I missed it, probably because I was busy writing those re:Invent blog posts, but I didn't see a lot of attention brought to this, and I think that's a mistake. You should definitely look at this. Now, let's look at a couple of resources I built over the last few weeks. It seems I have a bit of an automation obsession these days. I recorded two videos showing you how to deploy machine learning models. The first one is based on AWS Step Functions, where you can define state machines and plug in all kinds of services, including containers and SageMaker. This video will show you how to build that state machine to deploy a SageMaker model. The other one does something similar. I train a machine learning model locally on my Mac using XGBoost 2.0 and use an open-source project called MLflow to deploy that model locally and test it. Then, I deploy it to a SageMaker endpoint. This is a convenient library if you're working with various frameworks and want to deploy locally or in the cloud. In a few days, I'm also running a SageMaker Tech Talk. It's already recorded and will be broadcast on the 27th. I'll cover all the latest announcements for re:Invent, with a long demo and lots of code and new features. You can join for free, of course. I'll put all the URLs in the video description. That's it for this episode. Don't forget to subscribe to my channel to be notified of future videos. There are plenty more coming. Until next time, keep rocking.