Automated Machine Learning (AutoML) seems to be everywhere and every Analytics product and SaaS offering seems to have some element of AutoML built into them. Part of the reason for this is because most of the market analysts, such as Gartner etc., have been rating Machine Learning (ML) products and services based on them having an AutoML feature.
Some of the benefits of AutoML is it will automatically generate a ML model for you without you having to worry about any of the technical details and the various statistical tests to measure if the model is useful. This kind of message has resulted is lots and lots of articles talking about the death of the Data Scientist, as they are no longer needed. We must remember ML is only one of the tools and skills of the data scientist.
This can all sound great. No need to hire these expensive data scientists, I can just use this AutoML software to create a ML model, for my data, and life will be good with all these wonderful predictions. Just think of the money I’ll be making and saving!
Where the fun comes into all of this is when someone issues legal proceedings based on what one of these AutoML models has predicted. The AutoML has made an incorrect prediction. The problem you now face, probably in court, is trying to justify the prediction by saying the machine/computer/algorithm made it, and you have no idea how or what it is doing to make the prediction. Good luck in a court explaining that to a judge and/or jury. Be prepared to hand over lots of money
What is missing is the human in the loop, and in most cases this will be the data scientist or machine learning engineer (or someone else with a really cool job title). Part of their job is to evaluate lots of difference models for you data (remember they will create lots and lots of models and not just one!), determine (from experimentation) what algorithms work best with your data and problem, optimize these models and assess the impact of changing hyperparameters, look at how these ML models are behaving, are there any biases in the model or data, use a wide variety of statistic tests to assess the models, examine how the model works with different sub-parts of the data (customers), look at any potential legal and legislative issues not just in one geographic but across many disparate regions all of which have different legal requirements, etc.
As you can see there are many additional tasks beyond the ML steps needed to create, verify and select a ML to use. All of this is before you look at how it can be deployed in your production systems/architecture and building out you MLOps.
One importing characteristic of having the human in the loop is Explainability. Explainability of the process followed, what models were produced, the effect of tuning and opimizing, possible biases and mitigating steps, etc etc The list goes on and on. This the role of the data scientist and now it might look like a good idea to hire a good data scientist who understands all of this.
Taking a little step back, AutoML is kind of good cool feature/tool. A lot of the main steps of creating all those ML models, tuning them and evaluating them, etc can be very boring work. You do same steps for each model and do it all over again for the next, and so on for the tens or hundreds of models you will be creating. Most data scientists will have scripts in their toolbox (based from their experience) to automatically perform all of these steps and output the results. I mentioned the word experience in the last sentence. It can take a bit of time to build up to this. The AutoML products will do all of this automatically for you hence you don’t have to hire a data scientist to do it (see what I said above about this).
I mentioned above some of the challenges and the need to keep a human in the loop. AutoML can be seen as another tool to assist the data scientist and not to replace them. AutoML can be used to to help the data scientist work towards identifying what ML models to use. But this can be a bit of a challenge to do. It depends on what product or library you use. Some AutoML solutions act as a black box. Kind of like the image at the top of this post. These are simple to use but the draw back is there is not explainability or ability of the data scientist to really assess what is happening at each step. There are AutoML products/solutions that allow you to inspect and monitor what is happening at each step within AutoML. The diagram given able is one example of this. This allows for the human in the loop and allows for explainability. If the data scientist sees some unusual direction being taken by AutoML they can see where and why this is happening and can take corrective action. AutoML isn’t a black box in this scenario.
I mentioned above, AutoML can be another tool for the data scientist to use. Look on AutoML as quick way to see what might be possible. Using the information from each step of AutoML, the data scientist can use this information to guide them towards creating a more suitable and usable ML model, and do so in perhaps a slightly shorter space of time.
Going back to the title of the post ‘AutoML, what is it good for?’, the answer really is ‘It Depends!’, but if you do use it, be careful how you use the models and results beyond doing some simple investigation. And be careful of product offerings saying you don’t need anything else.