“I’d like to use machine learning, but I can’t invest much time, or I don’t know how” – That is something you hear all too often in industry and from researchers in other disciplines.
In recent years, the demand for machine learning experts has outpaced the supply, despite the surge of people entering the field. To address this gap, there have been significant strides in developing user-friendly machine learning software that can be used by non-experts.
AI That Builds AI
The resulting demand for hands-free solutions to machine learning has recently given rise to the field of automated machine learning (AutoML)
The Problem with Machine Learning
The authors of ‘Automated Machine Learning’1 explain it best – The past decade has seen an explosion of machine learning research and applications; especially, deep learning methods have enabled key advances in many application domains, such as computer vision, speech processing, and game playing.
However, the performance of many machine learning methods is pretty sensitive to design decisions. Since this requires both expert knowledge and some intuition that comes with experience, this becomes a considerable barrier for new users. This is particularly true in the booming field of deep learning.
Here human engineers need to make many design decisions like choosing suitable neural architectures, training procedures, regularization methods, and hyperparameters of all of these components. This all is necessary to make the networks do what they are supposed to do with adequate performance. This process has to be repeated for every application. Even experts are often left with tedious episodes of trial and error until they identify a good set of choices for a particular dataset.
Source: AutoML Course: Chapter 1 – AutoML.org
Why Use AutoML?
In a typical Machine Learning workflow, Data scientists indirectly teach the machine how to learn. This is often an iterative and tedious process that requires expertise, especially in deep learning. The design decisions involved are often not intuitive and error-prone, even for experts. Even after putting in all this effort, it is often challenging to achieve State Of the Art results.
The field of automated machine learning (AutoML) aims to make these decisions in a data-driven, objective, and automated way. This helps to reduce the complexity of our task. The user simply provides data, and the AutoML system automatically determines the approach that performs best for this particular application. Super Easy, right?
AutoML makes state-of-the-art machine learning approaches accessible to domain scientists interested in applying machine learning but do not have the resources to learn about the technologies behind it in detail. This can be seen as a democratization of machine learning. With AutoML, customized state-of-the-art machine learning can be at everyone’s fingertips.
So do we no longer need Data Scientists?
AutoML seems like an almost magical tool that can enable anyone to be a Machine Learning Guru. It would appear to be cheaper and more effective than hiring a machine learning expert. But that’s not actually true. As powerful as AutoML is, it has a lot of drawbacks as well. AutoML is still an AI in itself, so it suffers from the same weaknesses. It is susceptible to biases, which need to be handled by experts.
It also becomes difficult to understand why the overall performance is terrible and how to fix it. Using AutoML without understanding can result in the deployment of inaccurate and unfair models. There are a lot of jobs done by data scientists that AutoML just can’t accomplish.
It is more likely that AutoML and experts work together as a team, each doing what they are good at. This would allow scientists to use their time more efficiently and give better results faster.
AutoML still has some kinks that need to be worked through. But even today, it is mature enough to help both experts and non-experts alike. It is a big step towards our goal of Democratizing AI.