We'll continue our tour of pre-built models with high performing custom ML models with one surprising twist. You can build them with no code. Welcome to AutoML. Let's say I'm a meteorologist and I want to predict weather trends and flight plans from images. There are more than 10 different types of clouds, and each type could represent different patterns of weather that I may want to know ahead of time before it strikes. If I'm a pilot, or farmer, or a travel agency, or even just someone who wants to know if it's going to rain today or not. If we want to predict on weather, this means we need to identify not just that there's a cloud, but also what type of cloud. For example, a cirrus cloud is usually associated with fair weather, whereas cumulonimbus cloud usually foreshadows rain. So let's revisit our Vision API and see how well it does on this new problem. Here we show the image to the Vision API. After I uploaded the image to the Vision API, here are the results for the labels it inferred. The pre-trained model likely was never taught to recognize cloud types to this granularity. We need something a little more custom that we can train ourselves. If you need to extend the capabilities of the AI building blocks, consider AutoML. Here we will use our own labeled dataset of cloud images and train a custom model using AutoML Vision. Let's take a look at the AutoML Vision in action. In this dataset, you've got five image classes corresponding to five different types of clouds. Click on the different labels to filter by cloud type. The numbers tell us how many images are associated with each particular label. Once you've uploaded and labeled your images, you're ready to train your model.

You can specify your model name and your training budget. You actually get one compute hour for free for training. But for datasets with more than 1000 images, you can train for longer. So let's imagine that we received an email notification indicating that our model has been trained and now we want to evaluate it and look at its performance. Simply click on "Evaluate" up at the top and it's going to give you a lot of data about your model. You'll notice a few key metrics that are included here. Area under the precision-recall curve, and you can think of precision as a measure of the quality and recall as a measure of the quantity. So precision is the number of photos correctly classified as a particular label divided by the total number of photos classified with that label. So for all photos that our model classified as cumulus what percentage were correctly classified? Recall is number of photos classified as a particular label divided by the total number of photos with that label. So what percentage of my cumulus labeled photos were correctly classified as cumulus by my model? So you want to think about for your use case, if you care more about classifying photos correctly precision or getting a high number of correctly classified photos. Recall. You probably cared about both, but the metric is typically plotted as a trade-off. Ideally, you'll want to model in the top right corner. Now another great feature on the evaluation page of AutoML Vision is this confusion matrix. So you can see here the photos and which ones were labeled with a predicted label that was either true or false. What you want ideally is a diagonal line like we see here. But it looks like our cirrus predictions aren't performing as well as the other labels. So you can actually click on the items in the confusion matrix to see the photos that were misclassified. So here if I click on this box, it shows me that 23 percent of the photos for cirrus were actually classified as altostratus. So I'll click on that here and it's actually going to bring up the specific images that were classified correctly as cirrus, as well as the ones that are classified incorrectly as altocumulus.

Lastly, let's show a prediction using this particular ML model. So you can navigate to the "Predict" tab and upload an image. So I'm going to find an image on my computer of a cloud that I want to have the computer identify and predict its cloud type.

So you can see I've uploaded my photo here and it's predicting it's a cumulus cloud which is correct, and there you have it. That's the AutoML Vision product in action. The big takeaway here is not what we did, but what we didn't do. As we talked about at the beginning, typically when building a custom model, you have several steps which are complex and time-intensive. With AutoML, we don't have to do any of that. However, as great as AutoML is, it can't solve every ML problem, which is why you learned how to build a CNN by hand earlier in the course. Behind the scenes, AutoML is powered by the latest ML research. While your model trains, the AutoML platform actually trains and evaluates multiple models and compares them against each other. This Naznet approach or neural architecture search produces an ensemble of ML models and chooses the best one. Like go and self-driving cars, deep learning is now doing. Deep learning as you saw with the neural architecture search. But how well does a code list model you build with AutoML compare with some of the other image classification models you may have heard of? This graph is a refresh of the best models for ImageNet published in 2017. The x-axis is model size. The y-axis is accuracy. The best possible model position would be the green dot. What you want is a small model left on the x with great accuracy, top of the y. Our world has been trending towards big heavy models that aren't exactly like a brain. That's the black line. AutoML powered by Naznet is shown in the red line. The AutoML networks are smaller and more efficient. I'll provide a link to the Google AI blog so you can track the latest developments. Here's a final recap on when you should use the Vision API versus AutoML Vision. Recall that with AutoML Vision, you are primarily doing classification on image data that you are providing to the service. We did this in our demo where we needed the model to learn a label, the type of cloud that the Vision API didn't know. AutoML requires no coding experience, while the Vision API requires you to be familiar with invoking APIs with a programming language of your choice.