Automation has become a common buzzword in the ongoing conversation about artificial intelligence, as software shows potential to take over the work of accountants, factory workers, writers and even therapists. Now, AI is even beginning to automate itself in a process known as automated machine learning.
Automated machine learning, or autoML for short, essentially has algorithms take over the process of building a machine learning model. It handles the more mundane, repetitive tasks of machine learning, with the promise of both speeding up the AI development process as well as making the technology more accessible.
What Is Automated Machine Learning (AutoML)?
In recent years, there’s been a surge of interest in autoML’s potential to simplify the otherwise complex world of machine learning. DataRobot is often credited as one of the first companies to bring it into public consciousness back in 2013. Since then, Meta has dubbed autoML the “backbone” of its AI, and Salesforce acquired data analytics startup BeyondCore to create its own Einstein AutoML Library. Meanwhile, major tech behemoths like Google, Microsoft and Amazon have rolled out their own low-code machine learning tools that utilize autoML techniques.
Such widespread industry adoption is significant considering the expertise needed to build cutting-edge AI systems is in such short supply — even at companies like these.
“To me, I don’t see another way forward except for these more automated approaches,” Sarah Aerni, a VP of machine learning and engineering at Salesforce, told Built In. “There are too many opportunities for AI and simply not enough people to onboard to the business, onboard to the tech, deploy it into production, monitor it, and continue iterating on it. To me, autoML is where that enters as a solution to scaling.”
Although the concept of automated machine learning has been around for nearly a decade, it remains a work in progress. If and when AI-made AI does reach its full potential, it could be applied beyond the borders of tech companies, changing the game in spaces like healthcare, finance and education.
“Practically anybody who uses machine learning will also use automated machine learning,” Lars Kotthoff, an assistant professor and researcher at the University of Wyoming’s computer science department, told Built In. “Eventually, this will really be deployed everywhere machine learning and AI is used.”
The Goal of AutoML
At first blush, automated machine learning may seem a bit redundant. After all, machine learning is already about automating the process of identifying patterns in data to make predictions. The process, which relies on algorithms and statistical models, doesn’t require consistent, or explicit programming. Once a machine learning model is built, it can then be further optimized through trial and error and feedback, meaning the machine can learn by experience and increased exposure to data — much like humans do.
What Is the Goal of Automated Machine Learning?
In practice, much of the work required to make a machine learning model is rather laborious, and requires data scientists to make a lot of different decisions. They have to decide how many layers to include in neural networks, what weights to give inputs at each node, which algorithms to use, and more. It’s a big job, and it requires a lot of specialized skill and intuition to do it properly.
The more complex the model, the more complex the work. And some experts say automating some of that work will be necessary as AI systems become more complex. So, autoML aims to eliminate the guesswork for humans by taking over the decisions data scientists and researchers currently have to make while designing their machine learning models.
AutoML: Machine Learning for Everyone?
Eventually, the goal is to get to the point where a person can ask a question of their data, apply an autoML tool to it, and receive the result they are looking for without needing overly technical skills. And while there are a growing number of companies looking to democratize machine learning through autoML, this technology is largely exclusive to people with AI and data science expertise. It’s a tool, not a specific platform; and it’s a tool with fairly narrow uses, according to Kjell Carlsson, the head of data science strategy and evangelism at Domino Data Lab.
Carlsson advises customers on ways they can scale their data science strategy and utilize AI more effectively, and he describes autoML as a sort of “booster” or “accelerator” for data scientists.
“It can make it faster for them to discover the features that they want to use. It can enable them to more rapidly narrow down which algorithms they want to use. And they can be helpful in, early on, identifying some problems with your data,” he told Built In. “It can be very useful for the proof-of-concept phase — to figure out, ‘Is this doable?’”
How Does AutoML Work?
Automated machine learning, Carlsson said, is “mostly about” supervised machine learning, meaning it gives users information about the outcome that they’re trying to predict by creating a model that identifies patterns in labeled data.
There are many types of machine learning, but with supervised learning, tagged input and output data is constantly fed into human-trained systems, offering predictions with increasing accuracy after each new data set is fed into the system.
For example, if a company wants to be able to predict whether or not somebody is going to buy its product, they first have to have a data set of past customers, organized by who bought and didn’t buy. Then it has to be able to use that data set to predict what a whole new set of customers will decide to do. Or, if you want a computer to be able to identify a cat in a video, you have to first train it by showing it other videos with cats so it is able to accurately identify one in a video it hasn’t seen before.
AutoML and Supervised Learning
Automated machine learning automates the selection of different variables in a given data set that should be used in a model, as well as the algorithms needed to create that model.
In the case of predicting whether a person will buy or not, autoML would be used to parse through the thousands of data points the company has on that person, and decide what pieces of information should be used in making an accurate prediction. It also automates the selection itself, and decides which model makes the most sense. This could be a logistic regression model, a random forest model, some sort of ensemble model, and so on — whatever is most applicable to the business use case.
Because autoML algorithms operate at a level of abstraction above the underlying machine learning models, relying only on the outputs of those models as guides, they can also be applied to pre-trained models to gain fresh insights without having to repeat existing research or waste computation power.
How Long Does AutoML Take?
Exactly how long autoML takes depends entirely on the amount of data being fed into the model, as well as how many different types of models are being applied. For standard, structured data sets (like customer data in a CRM, for example), Carlsson said it can be “super quick” to run an autoML model — as little as just a few seconds. In larger data sets, where the user wants to try out lots of different model permutations of different algorithms to use, it could take days or even weeks.
Popular AutoML Tools
So what autoML tools are available? These are just a few popular choices being used among business professionals to automate machine learning processes.
Popular AutoML Tools to Know
- Google Cloud AutoML
Aible’s suite of AI solutions works to automate data science and data engineering tasks across multiple industries. Its products can detect key data relationships, assess data readiness for model input plus augment data analytics and recommendations. Aible connects directly to the cloud for data security, and can be integrated with other tools like Salesforce and Tableau.
AutoKeras is an open-source library and autoML tool based on Keras, a Python machine learning API. The tool can automate classification and regression tasks in deep learning models for images, text and structured data. AutoKeras largely applies neural architecture search to optimize code writing, machine learning algorithm selection and pipeline design.
Auto-PyTorch, based from the PyTorch machine learning library in Python, allows for fully automated deep learning (autoDL) tasks. It automates algorithm selection and hyperparameter tuning for deep neural network architectures, and can support tabular and time series datasets. Auto-PyTorch applies Bayesian optimization, meta-learning and ensemble construction for automation.
Auto-Sklearn is an open-source autoML tool built on the scikit-learn machine learning library in Python. The tool automates supervised machine learning pipeline creation and can be used as a drop-in replacement for scikit-learn classifiers in Python. Like Auto-PyTorch, Auto-Sklearn utilizes meta-learning, ensemble learning and Bayesian optimization to automatically search for learning algorithms when given a new dataset.
Google Cloud AutoML
Google Cloud AutoML is a suite of autoML tools developed by Google that can be used to create custom machine learning models. Leading the suite is Vertex AI, a platform where models can be built for objectives like classification, regression, and forecasting in image, video, text and tabular data. Vertex AI offers pre-trained APIs and supports all open-source machine learning frameworks, including PyTorch, TensorFlow and scikit-learn.
AutoML Examples: Using AutoML in the Wild
AutoML can be used on advanced artificial intelligence applications, or simple problems often found in conventional businesses that simply don’t have the humans to do it all.
AutoML Use Cases
- Predicting customer churn.
- Building customer lifetime value models.
- Predicting equipment failures.
- Grouping like products together on an e-commerce site.
- Predicting the success of an email marketing campaign.
AutoML Example: Salesforce
Salesforce has thousands of customers that are looking to predict a variety of things, from customer churn to email marketing click throughs to equipment failures. And all of this requires lots of rich data that is unique to their specific business, which can be used to build customized machine learning models. Salesforce is focused on making the creation of these models easy and accessible to everyone through automated machine learning.
“In order to leverage that data,” Aerni explained, “[Salesforce is] not able to look at it. So we need to use automated machine learning approaches to train on that customer’s data set, in order to transform that data.” This extends into various stages of the machine learning process, from data preparation to training and selecting models and algorithms that are most appropriate — all of which is done in a completely automated way.
AutoML Example: Aible
Arijit Sengupta, the founder BeyondCore, which became Salesforce’s Einstein after it was acquired, created a company called Aible, with the goal of helping anyone build an AI model that creates value. His goal: “How do we empower everyone to be able to extract value from their data using AI?”
Aible does this by offering a suite of software. One tool focuses on augmented data engineering, another is augmented analytics, providing companies with key insights into their data in language they can understand. And a third offering is augmented data science and machine learning, where it handles the predictive model building while also factoring in all the benefits of correct predictions and cost of incorrect predictions.
For example, “What’s the benefit of correctly telling you that somebody will buy? What’s the cost of incorrectly telling you somebody will buy when they wouldn’t? And how much capacity do you have to pursue these prospects?” he explained. “And then our system automatically generates an AI or a set of AI that would create the most economic value, given your unique business.”
The Challenges of AutoML
Sangupta said he wants to distance the company from what is traditionally thought of as automated machine learning, mainly because he considers traditional autoML to be “completely useless.”
Challenges of AutoML
- Can’t understand the business context of the problem it is trying to solve.
- No standard for what a “good” model looks like.
- Doesn’t offer the “why” of its decision-making process.
- Too complex for non-data scientists to pull off successfully.
- Can’t automate ethics or fairness.
Misses Business and Human Context
“The problem with traditional autoML is that it doesn’t start from the business reality,” Sangupta said. “It just tries various parameters and a bunch of models, and comes back and tells you ‘Here is the best model.’ And that genuinely is completely useless.”
Here’s an example: Imagine the benefit of a sale at your company is $100, and the cost of pursuing a lead is $1. You might be okay with relying on a machine learning model that gives you 99 wrong predictions for every one person that buys $100 worth of product. But, then let’s say your sales capacity only permits 20 calls. That creates a whole new set of restrictions.
“The problem with autoML is that it is so powerful that if you don’t think through what you’re trying to do, and if you can’t explain the impact of that model on the business that you’re trying to understand in language that business users can understand,” Sangupta said, “you can really hurt yourself.”
Acts Like a Solution Looking for a Problem
That is perhaps automated machine learning’s biggest shortcoming: Its lack of business intuition. AutoML will certainly produce a production-ready model more quickly, but it won’t necessarily tell a user why they should use a particular model or what the business justification is — let alone offer a justifiable problem to solve amid a massive set of data.
“It’s been a solution looking for a problem for quite a long time,” Domino Data Lab’s Carlsson said.
Another issue that comes up is that there’s no set standard for what a “good” AI model looks like. Is it based on just accuracy? Does speed contribute? Or its ability to learn? Either way, Carlsson said those metrics very rarely match up to what the business problem actually is.
“The joke is that all of us can create a model that will predict terrorist activity with 99.99 percent accuracy — we just predict that there’s never any terrorism,” he said. “Terrorism happens so infrequently that if I just predict that terrorism never happens, I’ve got this super accurate model. But it’s a useless model.” Meanwhile, if you created a model that predicted whether or not a player should take another card in a game of blackjack with 50.1 percent accuracy, “that would make me phenomenally rich,” he added.
In short: It’s all relative, and autoML models can’t tell whether its own predictions are useless or not. And no matter how complex automated machine learning is, it doesn’t offer the “why” of its decision-making process, which is something most of us crave when it comes to trust.
But University of Wyoming’s Kotthoff said it is “quite challenging” to actually achieve that, especially in the case of autoML, “because of the complexity of this whole machinery and the many decisions that are being made automatically under the hood.”
AutoML doesn’t automate ethics either. There is no built-in conception of fairness. You can impose different constraints in an effort to be fair — like equal rejection rate, equal acceptance rate, equal likelihood of success — and then make sure that the AI serves that definition of fairness, but Sangupta says that falls outside the scope of what autoML is capable of doing because humans have to set those constraints.
“That’s the danger with autoML is you end up doing the wrong business things and you do the wrong ethical things because the only thing the autoML system understands is the data,” he said.
The Advantages of AutoML
All that being said, there are certain problems that are really well suited to automated machine learning. These are the problems that require the creation of hundreds of thousands of models, and then updating those hundreds of thousands of models on a regular basis.
Advantages of AutoML
- Can discover and train multiple models at once.
- Remembers and organizes extensive data.
- Reduces time, resources and monotonous workloads.
Tackles Varied Models and Complex Tasks
More often, these models mentioned are classified as forecasting models, Carlsson said. For example, if a healthcare provider wanted to predict demand for different units across their network of hospitals, they would need to not only create different models for each hospital, but also the different units within those hospitals, as well as different time frames (one week out, three months out, and so on). In the end, you end up with thousands of models, the creation and re-training of which requires an immense amount of work for a human data scientist.
“AutoML models work really, really well in these kinds of instances,” Carlsson said.
Greater Accuracy Than Humans
And autoML generally isn’t prone to the same kind of forgetfulness or shortsightedness that we humans are — especially when faced with big, complex problems.
“Using these automated approaches tends to get better results than humans can achieve, simply because the machine doesn’t make mistakes. It takes all of this information I gathered in a principled fashion and then makes the decisions based on that, where humans are prone to forget things,” Kotthoff said.
Saves Time and Resources
But, of course, the biggest advantage of automated machine learning is that data scientists don’t have to do the hard, monotonous work of building ML models manually anymore, he added. “It’s really something that, in the end, will enable humans to work better and do more work in a small amount of time because they don’t have to do the tedious parts.”
Will AutoML Replace Data Scientists?
Like all aspects of automation, autoML is not immune to the ongoing speculation of it replacing human employees, particularly those working as data scientists. Indeed, the “democratization of data science” was the buzz-phrase when DataRobot first brought this technology to public attention, and it has been reiterated by everyone from Salesforce to Google. But the idea of a business being able to use this technology with absolutely no assistance from data scientists whatsoever hasn’t quite panned out, according to Carlsson.
“Because people don’t know what data scientists do, there is this view of ‘Well, if we have the right tools then everybody will be able to do this and we won’t need data scientists anymore.’ I have really never seen that be true,” he said, adding that, if anything, he’s seen folks move in the opposite direction. Companies are hiring more data scientists. And training more data analysts so they can become data scientists.
In fact, Carlsson says, not only will autoML not replace data scientists, but data scientists are really the only people who benefit from this technology at all. And even then it’s only “incrementally beneficial” to them, mainly because they require so much additional guidance.
“Using these automated approaches tends to get better results than humans can achieve, simply because the machine doesn’t make mistakes.”
“When you’re doing the world of data scientists, the actual creation of the model is just one small part of this,” Carlsson said. Data teams might use autoML a little in the beginning to do some exploratory analysis, but when it comes down to making the “real model,” he added, they’re going to create it from scratch themselves. “It turns out, you actually need folks who understand the data, know how to look at and analyze the distribution of that data, and know how to analyze the results of that data — the validation of the data — in order for you to create a model that actually makes any sense.”
And Aible founder Sangupta says the folks who are worried about autoML replacing data scientists outright are missing the point altogether. He doesn’t think giving everyone the ability to build an AI model that creates value means we have to get rid of data scientists at all. Instead, he likens what Aible does to what the Netscape browser did for widespread internet adoption in the 1990s — it made this foreign and incredibly complex new world more accessible to everyday people.
“Every technology goes through this phase where, initially, you have these experts and only the experts can do it. But the real potential comes when everyone is empowered to leverage it. That’s what’s going to happen with AI. It has to happen,” Sangupta said. Otherwise, the power disparity between the “AI have and have-nots” will continue to grow.
Indeed, what artificial intelligence is capable of now is vastly different from what it was even just a few years ago, and it has had huge implications on how businesses are run. Ordinary chatbots are beating the Turing Test, AI is keeping pace with increasingly sophisticated cybercrime, and sales teams are working with more precision and information than ever.
“Our world is changing so fast that, without AI, you can’t compete,” Sangupta said. “When the internet revolution came about, a lot of companies that didn’t get on board died out. I actually think the AI revolution is going to be far more disruptive than the internet revolution ever was.”