What Will Machine Learning Do Next? (part III)

My last installment of this series speaks to what I believe machine learning holds for us in the future. Much of what I’ve talked about to date isn’t very far away, but here I look further ahead to envision some innovative and unexpected areas where ML may change our world.

Optimization

But first, in my last post I left out one area where ML is already having an impact. Its ability to handle massive complexity in dynamic environments will allow optimization. This certainly includes route planning and supply chain management and distribution – of tangible goods as well as intangible, like optimization of the power grid – Google claims saving up to 40% in energy usage in their data centers, and are looking to apply the same idea to the UK.

Then there are whole new areas to explore. Engineering has traditionally been a discipline that is exclusively the realm of humans. But using reinforcement learning, ML is now capable of designing parts that meet a particular function, but at lower cost or weight.

Whole New Areas

It’s been fun thinking about all the places where ML can improve or transform the existing ways of doing things. But of course, the most exciting things are the ones we haven’t thought of yet. I’ll try to think of a few.

Communication. It’s hard to imagine machines communicating – in a real sense – with humans, and even with each other. But ML can already converse using e-mail to schedule meetings. Over time, some human-human dialogue will be replaced by human-computer dialogue, or even computer-computer dialogue. ML has even enabled machines to create their own language to communicate and collaborate. Which raises a question: what happens if (or when) they create a language that we can’t understand?

But ML-enabled communication will help humans long before it replaces their role in the process. Imagine a Wikipedia that is tailored to you, your interests, and your understanding. When you ask a question, you get a response that answers the question with a level of detail that suits your interest (or your level education on the topic!), with an option to explore deeper. No more need to scroll through a lengthy article, or decide that it’s too technical, or decide it’s too superficial, and then spend time searching for another site that has the info you’re really looking for.

Machine competitions. As machines develop superhuman capabilities, there may be some areas where it will be interesting – and perhaps even entertaining – to see them compete. It won’t be interesting watching computers square off in a chess match (and indeed the moves may be too fast for us to follow) but I can see a market for robotic racing – whether self-driving cars around a challenging race track, or autonomous drones through an obstacle course in the next-generation Drone Racing League. Or ping-pong (warning: marketing hype, not real).

Entertainment and fashion. ML is already used in conjunction with human creativity to develop (or improve) TV shows and music, and to design apparel and footwear. Or DJ (although choosing tunes isn’t all that impressive). It’s not a stretch to say that, at least in some cases, soon the human will not be needed as part of the process.

Food and drink. ML is also being used to come up with new recipes, both for food and for mixed drinks. It hard for me to envision a market for this after the novelty wears off, but there’s no doubt that ML can break new ground by thinking out-of-the-(human)-box.

Research and Discovery. ML works really well when it can try things over and over, learn from failures, building on things that show promise. Imagine what can be done when ML creates its own models, too complex for humans to create or understand, and then uses those models to simulate alternative scenarios? Now the computer is the explorer, and can try out different solutions in its virtual world much faster than what’s possible in the physical world. Then its solution can be tested in the real world.

This approach could be the primary method of exploration and discovery in many scientific areas; it has high potential for medicine, materials science, chemistry, physics, and astronomy. In addition to benefiting the physical sciences, it could lead to new mathematical proofs, economic theories, and understanding of human behavior too.

War and Surveillance. This is a scary topic to think about but you can bet that militaries and governments around the world are experimenting with what ML can bring to the battlefield in wartime, and to the intelligence community in peacetime. War itself may not have a lot of patterns that ML can leverage, but with modern technology it is data-rich, and there is certainly opportunity for assisted analysis of intelligence information and for autonomous robots.

Art. There are many things that fall into the definition of art, but to the extent that what is popular can be interpreted through the lens of patterns, it’s only a matter of time until machines can create art. I don’t see them writing novels or plays any time soon, but certain kinds of painting, sculpture, and music that leverage variations on repeating patterns (even if those patterns are not perceptible to us) are well within reach.

Self-learning. Right now ML only learns what we teach it. It’s very limited in scope, and only works in narrow domains that don’t transfer to other areas. Eventually this may change, and ML may be able to expand its own use. In other words, it might be able to learn how to learn. Transfer learning is an area of research in deep learning that’s still very young, but has huge implications as it may solve the highly-specialized nature of ML, which is perhaps its single biggest limitation.

Final thoughts

No doubt, it’s an exciting future ahead. Did I miss anything? If so, please comment!

But before we think that everything will be taken over by machines, there is reason to temper the enthusiasm. In my next post I’ll explore ML’s weaknesses – and where it won’t be taking over anytime soon.