AutoML is a way for a non data scientist to create a custom machine learning model or deep learning model technically already in that first sentence I have introduced to new words machine learning model and deep learning model so what are these?
A machine learning model is something is a computer program that based on inputs given and trained on outputs a prediction based on real world or new inputs specifically though deep learning is a subset of machine learning deep learning is the use of a certain type of architecture that is used when creating a machine learning model some examples of these algorithms AR random Forest sliding image things like that
To do what is the proper definition of machine learning and deep-learning
So now that we know what a machine learning and deep-learning model is how does automl allow us to make one? well what Google does with this automl is they perform what is called neural architecture search what this is doing is they've basically created a program that starts training a model and record its progress it then starts training another model with a different architecture and it continues testing new architectures and comparing how each one is performing then once it figures out the best one that ends up serving you that one so it helps you figure you don't have to do any configuration and that is one thing it does
One way that I've heard it is that what Google's automl is doing is essentially peeling back the layers of an onion you think of a model has multiple layers that's what a deep learning model is right it's a model with multiple layers the beginning layers the ones that the core social at the bottom of the pyramid are the ones that provide kind of a general purpose in general idea to the model and then it's the layers near the end that are really starting to fine-tune what the model is doing and what it's looking at or what it's supposed to be responding with so Google's neural architecture search what what it's doing is it's trying to figure out how many layers do we need to pull back so that your model due to essentially make you the best model possible and they're able to do this with transfer learning and transfer learning allows them to take one of their general purpose models which is you know just has high accuracy across the board and they take that and they start peeling back the layers on the top and they they pull those off so they're going to take away some classes and and different things like that and then they start reapplying your new data in these layers instead and start training and starting so they have to figure out you know how many layers are we pulling off and how many layers do we put it on and what then is the best model to do that with so that's why Google automl takes some time and they have the note hours because they're running multiple experiments and they needed the more time they have you know the better simply the more variations they can try and they can determine hey is there a better way to tune this or two in that for your model see you
automl in the real world is basically just an API or a website that you can utilize to create these things currently automl Vision allows you to upload images and video with image type data you can do image classification of a single label which means your image has one thing in it so if you how are doing an animal classifier it would say does this image have a cat or a dog on it in it yes or no black or white the other thing you can do with images is multi-label classification so this image has both a cat and a dog in it or it has just a cat or just a dog this is a more complex model and will take longer time to train the other thing Google automl Vision can do is object detection object detection allows you to identify an object within a set of boundaries boxes so in this image this is where the dog is or this is the location of the cat object detection can also handle multiple objects in a single image. Automl Vision also has a video product it is currently in beta but it does allow you to upload video and track objects detect scenes and a few other options
Once you've decided what type of model you want to make all you have to do is give Google's automl images or video that is labeled. labeling images for image classification is not too bad labeling images for object detection takes a lot more time in training or labeling images for video I have yet to do cuz it is still in beta. one thing that I would like to look into is Google's image labeling service which allows you to trade a labeling task and submit it and get images labeled that would be awesome too have the ability to do. Google automl Vision I believe is a great tool that allows developers who want to incorporate machine learning into their applications businesses can also take advantage of allowing developers to implement these tools easily into projects now one thing is that automl is not going to replace a data scientist for example if Google's automl puts out a model and it dips below the threshold score that you wanted to achieve how do you increase it you don't have this experience in Google's tools currently don't help you with this so this is when you would need a machine-learning engineer to come in and overhaul your data and your model to see how they can improve it and essentially hopefully get you to a higher level than allow you to take over handling it again until it dips below your threshold and once again you bring in someone to do an overhaul