Hi everybody, this is Julien from Arcee. Welcome to episode 15 of my podcast. Don't forget to subscribe to be notified of future episodes. In these 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 AWS on machine learning, and we're going to look at high-level services, mostly Deep Composer, Transcribe Medical, and Personalize. So let's get started.
All right, so big news on Deep Composer this week. The service is now generally available, and we added a whole bunch of features. If you remember, Deep Composer was launched in preview at re:Invent 2019. You could use it in the console, either the virtual keyboard or a real keyboard, and record some tunes using pre-trained models to generate compositions. Now, you can all use Deep Composer. You don't need to have a keyboard; you can start with the virtual keyboard in the console. We added some features, including learning capsules—short lessons on generative AI and generative adversarial networks, introducing the theory, what they are, how to train, and how to evaluate them. This should give you confidence to move on to in-console training because now you can train your own models with Deep Composer. I'll show you the console in a second. We have easy steps to train a model without writing a line of machine learning code, just working in the console. Then you can use your own models to generate compositions.
The physical keyboard is now available for sale on Amazon.com. If you buy one and link the keyboard to your AWS account, you get three extra months of free trial. This is an incentive to buy the keyboard, but you don't need to buy it. You can use the virtual keyboard in the console or pretty much any MIDI keyboard out there. If you want extra time to train your models, buying the keyboard might be an interesting option.
If we move to the Deep Composer console, we can see a few more things. If you've never seen it before, here it is. You can play music, and we have some pre-recorded tunes that you can generate compositions for. Or, you can click on the virtual keyboard or play on a physical keyboard and record tunes. The main new feature is creating a model. You would go here, create a model, and select one of the two GAN architectures. We have details, and if you're not familiar with this, the learning capsules will explain what these things are and how they work. Then you select a dataset to train the model on—symphony, jazz, pop, rock. You can set some hyperparameters, but for the first few trainings, you can leave them as is. Just give the model a name and click to start. It will train for about eight hours, and then you can use that model to generate compositions. This is a similar experience to DeepRacer, where you train reinforcement learning models for the autonomous driving car. Just work in the console, tweak parameters a bit, and try to generate interesting models. This is now available to everyone, and you can have fun playing music. I'm not going to play anything because I literally cannot play the keyboard. But go and check out Deep Composer and play some music.
The second service I want to talk about is one of my favorites: Transcribe Medical. Transcribe Medical was also launched at re:Invent 2019 and is an extension of Transcribe, our speech-to-text service, for medical vocabulary. I've covered this service quite a lot, and I'll add links to the blog post and a short video demo I recorded at the time, showing me reading medical text, which I have no idea what it is, but the important bit is that Transcribe Medical picks it up perfectly and is quite impressive. The announcement here is that you can now do batch transcription. If you have a bunch of audio files with either medical conversations or medical dictation, you can just upload them to S3 and launch a transcription job. This is really easy: input your data in an S3 bucket and select either conversation type or dictation type. The difference is the number of speakers. For conversations, Transcribe will try to identify the different speakers, while dictation should only be one person speaking. Input data in S3, and output data in S3. You can encrypt results with KMS to keep medical documents safe. This is a cool new feature from Transcribe Medical, making it easier to process medical documents at scale.
The third service I want to talk about is another of my favorites: Amazon Personalize. Personalize is a high-level service that lets you build personalization and recommendation models easily. I've covered this quite a bit in the past. Now, we're happy to launch recommendation scores in Personalize. My colleague Brandon has a really nice blog post about this, and I'll include the link. Let me explain what this is. Previously, when you trained a recommendation model, the process was to upload a dataset to S3. The minimal dataset was a user-item interaction dataset, a CSV file showing that user123 has interacted with item456, etc., showing interactions between users and items, which could be anything—movies you like, songs you like, products you bought. We would train a model, which Personalize calls a solution, based on a recipe. A recipe is an algorithm for the specific problem, preprocessing steps for the data, and tuning steps to optimize the accuracy and machine learning performance of your model. You could either pick a recipe yourself or use AutoML to let Personalize figure out which recipe would work best.
Here's a model I trained a while ago using the HRNN algorithm, a deep-learning algorithm for recommendations based on recurrent neural networks. I trained it on this dataset and deployed it. Using the campaign, I could run predictions, enter a user ID, and get recommendations, such as a list of movies a user might like. However, I wouldn't get any other information besides the movies. Thanks to this new feature, we now have recommendation scores. These scores let you know which are the top items a user may enjoy. Each item in the dataset has its own score, and all the scores add up to one. If you have thousands of items, the scores will be tiny. The scores are relative to one another. Each item gets a score, and they all add up to one. If you're working with a ranking model, the scores will be higher because we generate scores only for the items that need to be ranked, a subset of the dataset. Machine learning enthusiasts will recognize the softmax function, which we use a lot in machine learning and deep learning. Each score is between zero and one, and they all add up to one.
The cool thing is, this is a model I trained ages ago, and I didn't have to retrain it. I just ran the demo again, and voilà, we get scores. If you have models already deployed and campaigns running, you don't need to do anything as long as you use one of the algorithms that support these scores, mostly HRNN and a few more. Just run predictions, and you will see scores popping up. This is one of the top requests I've received for Personalize, so it should make a lot of people happy.
Alright, that's it for Personalize. That's it for this episode. Again, please subscribe to my channel and stay safe. I hope to see you on the road pretty soon. Until then, keep rocking!