Apr 04 2019

Deep Learning is Changing the World

This is a big-picture story I have been following closely, partly because it feels like we are still on the steep part of the curve. Every now and then in the development of technology we hit upon something that may seem small at first, but in retrospect changed the world completely. Electricity is an obvious example. At first scientists didn’t think it would be useful for much, but once we realized we could power devices and do other things with electricity, we remade our world with technology that uses it. The digital revolution was another such transformative breakthrough. Such technologies don’t just win the game – they change the rules of the game.

It’s easy to see such disruptive technologies in retrospect, and much harder to anticipate them. We are somewhere between these two extremes with current technologies that are in the process of transforming our world. It is becoming increasingly clear, for example, that artificial intelligence (AI) is rapidly turning into such a transformative technology.

Specifically we are starting to see an explosion of deep learning algorithms. These are AI programs that can learn general rules or patterns from examples. They can be fed examples, or take in examples by observing the world, or even generate examples through trial and error. This technology also dovetails with the age of big data – we have massive data sets and now we also have the AI to make use of those data sets in novel ways.

As is often the case, the early use of a technology is for frivolous purposes, if only as a proof of concept. Deep learning algorithms have been demonstrated to the world by winning at Jeopardy, chess, and now the game, Go. The news of these milestones were a flash in the social media pan, commanding our attention for five minutes before we went onto the next thing. But I think in retrospect we will see them as the harbingers of a new age of technology. The age of AI.

One of the obvious examples of this is self-driving cars. This is a fairly tangible way in which the technology will directly affect how we live our lives. The indirect effects, however, are likely to be greater. One way in which deep learning AI may indirectly affect our lives in through research. A recent example shows the potential power of this approach.

Material scientists are researching a new class of substances referred to as 2-dimensional materials. The poster child for this is graphene – a one-molecule thick layer of carbon atoms in a “chicken wire” configuration. This material is incredibly strong and has other potentially useful properties. It can be rolled up into nanotubes, for example, potentially making crazy strong fibers or cables.

But this is still a new technology. Again – it has incredible promise, but we need to figure out how to manufacture it, to minimize flaws, and to develop variations with specific properties. The faster research progresses in this area, the faster we will see applications.

Researchers at the University of Missouri in the College of Engineering are using deep learning to fast track this research. They fed many known configurations of graphene, essentially made with other elements in the matrix, and the properties of those variants into a deep learning algorithm. The AI was then able to learn how the atomic structure relates to the ultimate properties.

With a few thousand examples and just two days of computation, the AI was able to learn and predict the probable properties of billions of other configurations. In two days it accomplished decades of research.

This will at least create the possibility of sifting through those billions of possibilities for the configuration that has the properties you want, and then to actually make the graphene variety and test its properties. This will be ultimate confirmation that the AI approach worked. This also has the potential of shaving years or even decades off the development of applications for graphene.

And of course this is just one of countless examples. I imagine pharmaceutical companies will want to use a similar approach, to sift through billions of possible chemical structures looking for the next blockbuster drug, with minimal side effects and low toxicity. The already do this, but this may make the process orders of magnitude more efficient.

Recently an undergraduate student used AI to sift through public data from the Kepler dataset, and she found two new exoplanets that astronomers had not previously found in the data. This is a great example of the marriage of big data with AI.

Deep learning AI has the potential to give us control over the mountain of data our civilization is generating every day, or that our research is generating. This, in turn, has the potential to accelerate certain kinds of research by orders of magnitude. And here’s the thing – we are still in the very early days of deep learning. It’s hard to imagine that over the next 50 years or so this technology will not become critical to our infrastructure. It will likely become embedded in the background of almost everything we do, like electricity is today.

Of course this also creates the potential for abuse. Combing through big data with AI is also a great way to violate our privacy, and to take control of our lives. I am not paranoid about this, but we do need to think very carefully about the implications of this technology.

I would not be the first to point out that giving governments the ability to not only know intimate details of our daily lives, our preferences, behaviors, and habits, but also the ability to manipulate them, is more than a bit chilling. We only have to look at China to see where this can lead.

We have all by now had the experience of having frighteningly targeted ads pop up when visiting a web page. We chafe at a corporation using our data to target ads at us, which, let’s face it, is an attempt to alter our behavior. This is powerful. It doesn’t have to work all the time on everyone to have a significant effect on collective behavior. Now imagine this power in the hands of a political party, and ideological group, or the government itself.

While I think our democracy is robust, I am deeply concerned about the fragility that has also been revealed. Get enough people to buy into your narrative, and anything is possible. Our democracy is only as strong as the people in it.

At the very least we would not slide blindly into a future where the dark side of big data and AI is used to subvert our democracy. We need to think carefully about what steps will help protect it.

Meanwhile, the potential upside is huge. When I imagine our future I see incredible technology, smart cars, clean energy, personal supercomputers, better medicine, and much more. I just hope we also have a functioning democracy free from AI facilitated autocracy.

No responses yet