Fear in the Age of Machine Learning
This is my third post in a series on artificial intelligence and big data. Having recently completed a PhD on the subject, I plan to cover several controversial topics related to AI and the work that I do implementing this technology for companies. Not to worry, no NDAs are being violated and no company secrets are leaking out through these posts.
This particular article is about uncertainty. It is about admitting and embracing fear in the face of machine learning dominance over technology companies.
What we know < What we don’t
The pessimist says “everything’s terrible, it can’t get any worse” and the optimist says “oh yes it can.” (West Wing S05E21)
I’m an optimist about pessimism. I have to be. Most startups fail. In Canada we have a catchphrase for ‘regular’ sized companies: Small and Medium-sized Enterprises (SMEs). These SMEs typically have fewer than 100 employees. Lean stage gate and other fail-fast methodologies are already a challenge for these SMEs trying to stay alive. There was already enough to be afraid of when taking the plunge to start or grow a small tech business. For example, access to seed funding is not easy to come by. Adding Machine Learning (ML) to the mix really disrupts the plans of many early stage companies that now need to know about AI in addition to the standard cake baking recipe for success (cloud, SaaS, REST, etc. etc.)
Never mind understanding how it works, many of these SMEs simply do not understand what machine learning is good for, while their stakeholders demand to know how they will defend against the oncoming army of AI machines.
Pretending we have power helps us feel safe. It is cathartic. But things are going to get tough for SMEs because of the AI onslaught. Let’s talk about what these companies are facing with a 2 year time horizon.
The Tsunami Reaches the Shore
Demand for ML services far outstrips supply, and most buyers of ML services have a poor understanding of where it is best to apply ML within the organization. The coming years will see a lot of creative destruction, and acquisitions by the giants in tech to protect their oligopoly. I like the term FANG used by market analysts to talk about some of these giants (Facebook Amazon Netflix Google). If you are a typical 100 person firm, the dynamics are not in your favor. Highly qualified data scientists are hard to come by, as they are being shuffled en-mass through the doors of either startups promising big equity or the big giants offering wheelbarrows of cash. Here you sit, stuck in the middle with few good options for your SME. Sure, things will stabilize 2 years from now. More data scientists and engineers are rushing to fill the gap right now, and will come online at some point… But what about now?
There is a pernicious greed versus fear calculation even for the consultants like me, who drop into companies to fill this resource gap. The law of supply and demand means that the intersection of low supply and high demand drives up prices, and so the consultants also prefer to work with well established companies or funded startups. Academia offers few reprieves for the SME. Companies are reaching straight into universities to hire talent, and working with a university to do non-research core development does not bode well for a project. University projects can run longer than expected, and provide no success guarantees. If the project fails, you don’t get your money back. Students come and go. Things happen.
Having taken the perspectives of companies and the workers, let’s have a look at the technology side of things. Uncertainty plagues the mathematical machine learning world too. You may not have noticed something odd about how machine learning encodes uncertainty. Have a look at the following image from Adrian Rosebrock’s excellent pyimagesearch blog.
There is something off here. Something not quite right. Can you see it yet? If you are a scientist, or deal with uncertainty regularly, you will notice the predictions are scalar. There are no error bars. So predictions like “I see a soccer ball with 93% confidence” become absurd when applied to real life. For example, is the prediction “99% chance of rain” less credible when you find out the real chance of rain is “99±25% chance of rain” or maybe it was “99±1% chance of rain”? It is not all bad news though. There is some exciting recent work by Yarin Gal and others on fixing this problem in machine learning. See more here.
Uncertainty is the order of the day. It will not go away. It is here to stay. Companies, people, and even machine learning models, need to better quantify, tolerate, and plan for uncertainty as we all get used to the widespread adoption of ML.
Best of luck to us all!