I’ve worked as a data scientist for over a decade, much of that time spent on training, testing, and calibrating machine learning models to address specific business problems at various scales. I’ve done a lot of data engineering, and data product management, and a whole lot of plain-old, business-intelligence-style analytics, but more than anything else, I’ve done machine learning.

I’ve reached a conclusion: machine learning is a dead-end street. I don’t mean machine learning isn’t worth doing. Machine learning has been and will continue to be useful. Businesses that aren’t using it still have lots of ways they can benefit from it, but businesses that already use it will, I am convinced, see a plateau in those benefits very soon, if they haven’t already. And investors are starting to notice that.

That’s the cost of a dead-end: you reach a point where staying on the road stops bringing you closer to your destination. Machine learning does a lot of useful things, but not what most businesses really need.

Exciting math, boring results

Molly Sauter wrote about machine learning as “memory” systems, of which she distinguishes three:

  • “Predictive text, those systems concealed within your phone’s keyboard that prod you to call your dad ‘pookie’ because that’s what you call your girlfriend;”
  • “Reminiscence databases like Facebook Memories or Timehop;”
  • “Data doppelgangers constructed for ad targeting, the ones responsible for those socks that follow you around the internet even after you’ve bought six pairs.”

So we have systems that predict the next step in a sequence based on other steps taken, that flag particular past steps in a sequence as being important in the present, and that recommend a next step based on past steps. Sauter concludes:

Each interacts differently with the data it collects, representing it to guide or nudge you according to different models. But the core of these models, their fundamental shared strategy, might be reduced to: “Those whose past is legible will be exhorted to repeat it.”

All machine learning assumes (1) the future will be like the past, and (2) what happened in the past is more-or-less what we want to happen in the future. Cory Doctorow summarizes Sauter’s argument as “machine learning is fundamentally conservative, and it hates change.”

It’s not common for executives or investors or anyone associated with a business to ask “how can we ensure a really conservative growth rate?” People are usually more ambitious than that, but machine learning isn’t. In many cases, there’s a fundamental misalignment between what the tool does and what the tool’s users really want to accomplish.

Over time, a business that hires a bunch of data scientists to support machine learning will see the incremental value those data scientists offer go down because there just aren’t that many things that need machine learning. Unless you’re one of the really big tech companies, rapidly and continuously introducing new products and capabilities, you reach the end of the road but keep the engine running. The value added by machine learning has plateaued, but sunk costs and lack of alternatives keep companies investing. That might be one of the reasons many companies have constant trouble keeping their data scientists: they run out of interesting things for their data scientists to work on.

Efficiency is more important than accuracy

I’m convinced that most people who talk about the value of machine learning - even those who actually do machine learning - miss its core value. Machine learning frees up humans’ time by automating things that computers can do roughly as well as humans can.

Most people know what it’s like to have 10 things on the to-do list but only enough time to really do 2 of them well. When you know you’re only going to get a chance to pay attention to - let alone make a decision about - 2 out of 10 important things, it gets difficult to decide what most deserves your attention. You end up delegating a few things (and those things often get re-delegated), and then choose the two things that get your attention based on how easy it is to do the work, or how familiar they are to you, or sometimes just on your gut feeling about what you should do. I’m not finding fault with that - it’s what humans have been doing for a very long time. We all do it.

But let’s say five of those 10 decisions are machine-learnable: you have historical data associating attributes with outcomes, and you’d be content to just maintain the outcomes you’ve seen previously - bonus points if you can see a modest increase. You might even be willing to see a modest decrease in those outcomes if it meant you only had five things fighting for your attention instead of 10. Machine learning can create a real, substantial impact on your business, and all it has to do is not be obviously a lot worse than having a bunch of overworked humans do it.

Machine learning adds even more value by making logging process details to make the process auditable. If a human process isn’t performing well, you have to commission a study or at least try to talk to everyone involved to figure out what’s wrong. If a machine-learning process isn’t performing well, a reasonably competent engineer can crack open the logs and find things to change. So in multiple ways, machine learning conserves human attention and effort.

These benefits don’t come from data science or machine learning. They come from engineering. As far as I can see, there is no hard-and-fast demarcation between engineering and data science, but I’ve found it convenient to draw the line between processes and decisions. If automating a process - say, moving data from one place to another - I find people usually want an engineer. If automating a decision - you have data in one place and you have to choose which one of three places it gets moved to - it sometimes helps to have a data scientist.

That division is disappearing. Companies like DataRobot have already productionized a lot of that decision automation. There’s still some distance to cover - in particular, I see those solutions fail for businesses who don’t know how to deal with missing data or high dimensionality - but those aren’t insurmountable obstacles. Increasingly, machine learning is becoming just another engineering problem. Engineering without machine learning solved a lot of business problems, but it didn’t solve them all, which is one of the reasons machine learning has been so widely adopted. Machine learning now is where engineering was before machine learning happened. It’s usefulness is very apparent, but also apparently tapering off.

An alternate route

Experimentation isn’t a dead-end street like machine learning is. I understand if you’re not willing to take that statement faith: maybe we’re just nearer to the limits of machine-learning’s usefulness than we are with experimentation. That’s a reasonable doubt. There are a few reasons I think experimentation gets businesses to where they want to go:

  • Experimentation makes history while machine learning simply repeats it. Experimentation is a data-generating process: any experiment produces information that did not exist before. That data can be structured, stockpiled, and mined the same way any other data can. And it’s data on real behavior - not opinion or intended action or extrapolations from a panel or aggregations from a geography. It’s actual people making actual decisions.
  • Experimentation is as disruptive as machine learning is conservative - it can’t make sense of anything it hasn’t seen before. That drastically limits its applications, regardless of how accurate the model is. (Incidentally, this also makes it very easy to disrupt what value machine learning can offer). Experimentation is the only way for a business to try something completely new in a data-driven way.
  • Experimentation deals directly with risk rather than just probability. Machine learning says “here is knowledge that may help us.” Maybe it gives you a score to tell you how confident the model is that that information will help. Experimentation is the only way to say “here is knowledge that definitely either did or did not help us.”

There is one thing machine learning has that experimentation clearly doesn’t: excessive computation and infrastructure costs. Particularly the most cutting-edge deep-learning variants of machine learning require specialized hardware and huge resources to work. Google’s AmoebaNet took 450 K40 GPUs 7 days (that’s 3150 GPU days) to train. Most companies don’t have those kinds of resources. What’s more, they don’t need them. Experimentation costs less and offers greater and more reliable value.

Experimentation is so much more than A/B tests. A/B tests are to experimentation what logistic regression is to machine learning: you can use it, and many people do, but it’s incredibly easy to use it wrong, for it to feed you garbage results. Experimentation is only really likely to give dependable value that doesn’t plateau when it is conditioned, continuous, and connected.

  • Conditioned. If you just split your users into two groups, assign a different treatment to each group, and compare the response between groups, your experiment will probably lie to you. You need to Identify important differences in your user set (something far more nuanced than segmentation), and assign treatments based on that variation. And then you need to use that conditioning to adjust your results. (See here for more details.)
  • Continuous. If you have to manage experiment logistics and in-flight quality control for every experiment, you’re not going to get much out of your experiments - it will just be too much of an attention suck. Likewise, if a human has to decide how to use experiment results or manage logistics and quality control, it’s going to require too much effort to be viable. Experimentation is just as much of an operational capability - perhaps more so - than it is a research capability. Past experiments need to be pipelined into the next experiments without needing a human in the loop.
  • Connected. Apply results from small experiments to your entire customer base: learn from a few, but make decisions about all. Then consolidate results across experiments into customer indices for wider value. Plug in a new message, offer, or creative and start generating intelligence about it automatically. Business problems need to be pipelined into experiments. This isn’t about generating knowledge. It’s about giving your business a new basis for taking action.

Move from machine learning to experimentation

Of course, businesses shouldn’t drop machine learning and pick up experimentation instead. Those aren’t mutually-exclusive options. In fact, machine learning can be enormously helpful in the interpretation and pipelining of experimental results. But value from machine learning tapers over time: if your business uses machine learning and you haven’t seen that tapering yet, you will.

This is one more example of something experimentation can do that machine learning just can’t, as Eric Ries points out:

Whenever you’re not sure what to do, try something small, at random, and see if that makes things a little bit better. If it does, keep doing more of that, and if it doesn’t, try something else random and start over. Imagine climbing a hill this way; it’d work with your eyes closed. Just keep seeking higher and higher terrain, and rotate a bit whenever you feel yourself going down. But what if you’re climbing a hill that is in front of a mountain? When you get to the top of the hill, there’s no small step you can take that will get you on the right path up the mountain. That’s the local maximum. All optimization techniques get stuck in this position.

Machine learning has a built-in capability gap. There is no bridge over it. Conditioned, continuous, and connected experimentation is the alternate route.