Hi, I'm Carolyn Ujcic. I lead a team of machine learning engineers who have successfully implemented many machine learning projects across various industries. So far in this course, you've been applying machine learning to your structured datasets with custom models. It's time to widen our aperture and look at unstructured datasets, like images and also other approaches to doing ML that don't involve custom model building. Here's our agenda. We'll first see how ML on unstructured datasets is driving value for businesses. Then, we'll look at the intuition and models behind unstructured data. After that, is a critical topic on choosing the right ML approach. Whether to build a model from scratch, or use pre-existing building blocks. Lastly, we'll do a demo to classify text using three different approaches and compare the results. Here are just a few customer examples of companies who have used the ML tools on Google Cloud Platform on their datasets. The left to right ordering is reflective of more low-level programming to more abstraction and UI based. Aucnet built their own custom model to classify images of car parts and estimate price. Ocado used parsed results from the Natural Language API to route customer emails to the correct responders. Giphy use the out-of- the box vision API to find the text and memes using optical character recognition. It then can reject inappropriate uploads based on sentiment or keywords. Uniqlo designed to shopping chat bot using the dialogue flow UI. Dialogue flow is a Google owned company which specializes in building ML based interfaces like intelligent chatbots. Let's look at some other use cases for ML and business. You might be Airbus and use machine-learning to differentiate between Clouds and snow cover. If you're stumped like I am, the Clouds are in the upper right part of the right image highlighted in red. You might be an economic forecasts firm looking to track the global fleet of container ships via satellite imagery. Knowing the amount of cargo being carried, might help improve your economic forecast by days or months ahead of the official numbers. Medical images are ripe for innovation. For example, you could diagnose medical conditions like diabetic retinopathy earlier when it's easier to treat and prevent blindness. The key takeaway is that ML can automate tasks that may save or assist a human team. Recently, the latest models are even outperforming humans in some domains. Also image classification as a field is more than just binary classification tools. Later in this course, you will experiment with the pre-built Vision API model which allows you to pass through a JSON request and get back a ranked list of associated labels for your image. Also if you have more than one subject matter in a photo, you can draw bounding boxes and classify pieces of an image as well. Modern image classification models can even generate captions describing what is going on in the image, like a map of dependencies, like two hockey players are fighting over a puck. Here it's important to call out that even the best models can and will make mistakes in their predictions. Like the road sign captioned as a refrigerator filled with lots of food and drinks. As you saw, there were a lot of impressive uses for machine learning these days. Like detecting objects and images, helping to detect diseases and even enabling cars to drive themselves. But AI can also be used in more playful ways too. Through a pose estimation model, a Google AI experiment called moved mirror, can match your real-time movements to hundreds of images of people doing similar poses from around the world. Feel free to try it out yourself and have some fun. Then tune back into this course to learn how image classification models extract features like these from images.
