AutoML is deemed as the next wave of machine learning as it allows organisations to deploy models and make informed decisions quickly. For years, AutoML was only used for tabular data, but now it has been enhanced for creating models that deal with text and images. Such enhancements lead to a rise in the adoption of AutoML solutions among AI-driven organisations. Today, firms are actively embracing automation in ML for transforming and their workflows to address customer demands.
Reports suggest that AI skills gap is causing hindrance in technology adoption. Thus, firms have veered towards adopting AutoML solutions as a workaround to the absence of relevant talents in the landscape.
How Effective Is AutoML
A trio of Google’s AI experts used AutoML in a Kaggle competition against top data scientists and finished second after leading most of the way. Such instances demonstrate the capability of AutoML solutions in the current landscape. However, that doesn’t mean a non-expert can give a run for their money to data scientists. In fact, it affirms the need for data scientists to leverage the tool to make the most out of data. Thus, addressing talents gap with AutoML cannot be the right solution.
Subscribe to our Newsletter
Join our editors every weekday evening as they steer you through the most significant news of the day, introduce you to fresh perspectives, and provide unexpected moments of joy
“Although these tools have simplified the training and evaluation of ML models, there is still a good amount of data science knowledge that is required for creating highest-quality models,” mentions Erin LeDell, chief machine learning scientists at H2O.ai, on the blog.
Diversity Of AutoML
Using AutoML non-experts and data scientists are able to find solutions to the ML problems due to an increase in the availability of various open-source and paid AutoML tools such as H2OAutoML, Auto-Keras, DataRobot, among others. The challenges with these solutions are that they can only handle specific tasks. “Majority of AutoML tools are designed for most common use-cases,” describes Erin. “A few support time-series data and raw text, while others are good with image and text.”
Since data comes from different sources and are in different formats, AutoML is not compatible with various datasets. Thus, it limits the capabilities of users. Only using for specific data can be helpful for a small business, but for large firms, it is ineffectual.
However, even with narrow used cases, AutoML is automating AI to assist firms in achieving their objectives. It has the potential to manage the complete ML workflows but needs to be further enhanced for broader ML activities.
AutoML Only Solves A Part Of The Data Science Problems
Data scientists carry out numerous tasks right from data wrangling to finding statistics and building ML models. But, AutoML mostly focuses on choosing and increasing the efficiency of models. Further, it does not have a say on the quality of input data, blindly takes and tries to improve the output by finding the right balance for parameters and hyperparameters.
Undoubtedly, it is allowing anyone with limited data science capabilities to deploy ML models and solve sophisticated business problems. But AutoML’s incapability to handle a wide range of ML practices causes hindrance as non-experts can only carry out limited ML tasks.
But, as per Kaggle ML and Data Science Survey, it was found out that firms are mostly working with relational, text, image, and video data, extensive data science expertise might not be required to do regular tasks.
Lacks Explainability And Featured Engineering
While AutoML is able to tune the hyperparameters, identify the best model for finding predictions, and provide other insights, it does not solve the problem of the black box. Instead, it has further created confusion about how it picks one model over the other.
Besides, there is no clear evidence about its efficiency while delivering the outcomes, any decisions made based on the derived conclusion can negatively impact organisations. There is a considerable gap between what organisations expect and what AutoML offers.
“One of the key ingredients for building great ML models has been often disregarded from AutoML systems: feature engineering,” wrote Marcia Oliveira, a senior data scientist at Skim Technologies. AutoML lacks human touch; data scientists gain superior data intuition with their experience, thereby can use their cognitive to assimilate the data at hand for moulding in a way that can result in desired outcomes. Developers often create derived columns from the raw data to add new variables for analysing on various aspects. Such techniques only work when data scientists find shortcoming while investigating data, but with AutoML, they fail to pinpoint the problems for optimising the data.
As a result, data scientists prefer to manually carry out feature engineering and use their intuition for obtaining meaningful insights into data.
AutoML is a wise step towards proliferating machine learning technologies, but we are way behind in completely replacing data scientists. However, companies like DataRobot are striving to eliminate the challenges in AutoML to manage the complete ML workflows. AutoML requires further enhancement to address the need of businesses to gain value out of it.
Although AutoML can still be leveraged to choose among different models for quickly making informed decisions, it might be difficult for firms to match the pace of the evolving data science domain with AutoML capabilities.