The next evolution of machine learning: Machine teaching

When trying to solve a Machine Learning problem, it's often not hard to set up an algorithm that does better than pure chance. As they imply in the video, and from my own (admittedly rather limited) experience, the hard part is selecting the correct algorithm for the job and tuning the parameters just right. Although this is mitigated by so-called "great estimators" like Neural Networks, achieving the highest accuracy requires both a Machine Learning background and a decent knowledge of the problem space (i.e. domain knowledge). Is over-fitting likely to be an issue? Is there a class imbalance? Are the classes undergoing concept drift? How costly are false positives and false negatives?

This is simply another example of a trend that has been driving computers, and technology in general, since at least the industrial revolution: abstraction. Although abstracting away the complex to present an easy-to-use and universally accessible interface has been going on since before computers were even invented, this specific instance strikes me as radically different. Microsoft could be enabling billions of people around the world to deploy their own personal pseudo-AIs. Trying to decide what stocks to invest in? Deploy an AI to crunch the numbers. Not sure whether that girl in your class likes you? Ask an AI to assess the situation. Want to confirm if a country has nuclear weapons? Analyse the media with an AI.

On the negative side I worry that too much abstraction could be wholly misleading. After all, there are still many people who take everything said in the media or on the internet at face value. What if the oracle is wrong in a life-or-death situation? What if it's right in a malicious application? In this age of big data and non-privacy, just how scary are the potential abuse cases for a project like this? Moreover, will these pseudo-AIs just become another source of lying statistics due to flawed assumptions and implementation?

At the same time I look forward to increased adoption of computer decision making. Implemented correctly, a computer algorithm comes closer to true impartiality than most humans ever could. They do not get tired or distracted. For example, I cannot wait for the day when cars are driven by computers.

I can speculate about the ethical and physical ramifications of personal AIs until the sun blows up, but I think what's more important to consider is what comes a few steps after this in the "evolution of Machine Learning".

What do we do when we reach the singularity?

We're getting closer and closer every year. I think it's time we move past having theoretical discussions about it and actually start implementing some preventive measures. Let's not be as unprepared for and passive towards this issue as we were towards climate change. EMP bombs anyone?