4 Main Reasons We Are Better With AutoML
Change is scary. I don’t have to tell you that. We’re afraid of the dark because it blocks our sight. We associate it with being vulnerable and unprotected. We say we’re afraid of the dark, when really, we’re afraid of change.
Tonight, I engaged the #DataEveryone community in a discussion about AutoML, a word that has become increasingly associated with change and the fear of it. Will it take our jobs? Will robots learn to think on their own? Are we already obsolete and don’t even know it? Well here’s what I do have to tell you:
Speed isn’t Everything.
There are still many aspects of machine learning that data scientists in the thick of it still feel are meant for humans, whether it’s something as bottom line as correlations and causality or as advanced as feature engineering.
“Feature engineering takes knowledge of the domain, creativity and the ability to evaluate the possible information gain versus the cost of operating the feature in question,” said @pip_alise. Her experience in the area thus far comes from her former work as a data scientist for Spirit Airlines. The airline deployed AutoML to automate pricing for one of the smaller ancillary products, which consumers ultimately neglected to purchase.
Even basic statistical calculations may require human hands. Distinguishing between causation and correlation will always need people to interpret data and make sure we are fully understanding what is going on,” said @asuka_hayasaka. “It’s especially important when confounding variables that might explain relationships outside of what information it is given.”
Overall, we have to carefully gauge whether there is a need for AutoML in a given situation before charging forward. It’s not a good idea just to find an excuse to try it even if it temporarily puts us behind our competition. Just like finding significance within a data set, sometimes the best takes a while to show up.
And Where Exactly Do We Show Up?
Some industries will inevitably be drawn towards automation more than others, so I asked the #DataEveryone community which industries could benefit the most. Rather than shouting out industries, the discussion turned more towards determining which factors make an industry most likely to adopt AutoML.
@rladiesnola suggested that “every industry will feel the waves but some more than others at first. It really depends on who has the biggest budget to invest in new technology and the bandwidth to on-board it.”
More than just budget, we have to consider which industries have access to the large data sets suitable for introducing AutoML. Without the resources to deploy the technology and make changes within departments, the opportunity cost of switching operations to AutoML may be too high. Jumping into a project with AutoML out of excitement to try it and stay relevant may result in losses financially and of the right opportunities. “Unfortunately the corporate world, especially in the US, does not wait,” explained @asuka_hayasaka. “If market competitors are using some form of AutoML, so are we. We don’t know the negative effects until its too late.”
That said, in time, @pip_alise believes that AutoML will have universal benefit: “I think it’s a great way to have some models going that may not be high-risk or requiring as complex feature engineering or processing.” One suggestion, from @asuka_hayasaka, was to use AutoML while hiring. “I would like to see it influence the consulting field and something like leadership training. Or perhaps selection, if we can move away from something like Facebook likes or Meyers Briggs to something more valid and reliable,” she said.
And speaking of reliable, it would be great if we could:
Stop Talking in Codes.
@asuka_hayasaka said it best: “ It worries me how fragmented conversations around AutoML are. That means someone with the power to harness AutoML may not have the knowledge of what it even is or how it can help.”
And the scarier truth is that if practitioners struggle to capture the essence of AutoML, translating it for those without machine learning knowledge will be difficult. That would undoubtedly go against the goals of key players in the data science industry. “I hope that giving non-data scientists the ability to transform data into information will improve their analysis and the conversation around a subject,” said @ogustovo_com.
The only way to up our collective knowledge as we discover AutoML’s capabilities is communication. Conversations may be overlooked by most data practitioners in favor of pushing calculations towards profit. However, if we make a point of engaging in discussions as we progress, true understanding about AutoML may be retained and used as a foundation for future growth.
I mean our growth as a species…
At the risk of sounding dramatic, I’ll address what I feel the introduction of AutoML means for the future of our kind. It’s most likely not even the death of our jobs; there’s evidence of that emerging even now. @asuka_hayasaka said, “I saw recently some post about not using self check-out because it was taking people’s jobs, and then I saw a counter argument from a cashier saying they need self check-out to be able to do other tasks because they are understaffed.” This dichotomy shows that automation, like any other feature, can be both harmful and helpful to us. @asuka_hayasaka clarifies that in the case that AutoML might be harmful, “those implementing it have a social responsibility to come up with different jobs for those they are replacing.”
This, of course, would require a lot of innovative thinking on our parts. And this is where the growth of our species comes in: We can’t be lazy. Applying Auto-ML may improve efficiency, but it can’t be about doing less work. If we find a way of transferring our current responsibilities to technology, we have to find new ones, otherwise we will shrink as a race. And we’re better than that.