Infinite U – Part 1 (or How Machine Learning Could Reshape Human Learning)

Co-authored by Mickey McManus and Randy Swearer

At this moment traditional educational institutions are facing an existential threat while industries are finding it harder and harder to fill whole job categories. And yet, we don’t think anyone now disagrees that the next round of automation will collapse, remake, or entirely shift broad swaths of jobs and even the meaning of a career. If the last wave of automation got rid of blue collar jobs the next round will be much more focused on the white-collar world and everything in between. These countervailing trends in education, technology, and industry make this the perfect time to consider a radical reboot of learning and ask what sort of “new-collar” roles will be resilient and necessary after the dust settles. 

The state of the art

The next few years will see a massive challenge in the form of resource allocation. It’s not that there won’t be jobs, but that the right person with the right skillset, mindset, or toolset won’t be in the right place at the right time. Our idea of what a job is will likely change as well. This shouldn’t be too surprising; our great grandparents might have a hard time figuring out what we do for a job today.

How do we learn new things? How might we accelerate not only learning but unlearning as job categories collapse and what we thought was secure becomes a fading mirage? What role will machines play as we shift from an education model–that is more akin to a factory pumping out one-size-fits-all workers ready for a single career in their life–to an era where learning is a verb over our entire lifetime? This is a shift from “K through 12” to “K through Gray.”

Automation has the potential to open entire new pathways and move us towards a golden era of education and collaboration. We could ultimately dream much bigger dreams for ourselves and our children if we can get from here to there. But the journey isn’t obvious and won’t be easy. There will be casualties along the way, false starts, and times when we can’t see the far shore. As a start let’s collect a few observations and experiments that hint at the beginning of a map to the road ahead.

Desirable Difficulties

Everyone has heard of the bell curve. The one below maps the distribution of a population in terms of where they fall from amateurs to virtuosos on a given toolset, mindset, or skillset. Everyone falls somewhere on this curve.

If you’d like to get better, you can practice with someone who is within your range but maybe a little less skilled than you. When you mentor someone you usually get the best part out of the deal because you must focus on how and why you do what you do (teaching something is a great way to figure out what you’re really doing). It’s sort of like playing against a younger version of yourself. Think Venus playing against Serena. They both get better by pushing each other up the skills curve.

It’s also good to challenge yourself by playing against a sort of future you. Reinforcing learning by playing a virtuoso teaches you what’s possible and pulls you over the curve. When done right you find a flow. The younger “you” sets the lower boundary and pushes. The future “you” sets the upper boundary and pulls. If it’s too hard you give up, too easy and you get bored. Cognitive neuroscientists like Adam Gazzaley who have pioneered ground-breaking new science in neuro-plasticity calls this the “desirable difficulties” zone.  

For more on the idea of “flow”see Mihály Csíkszentmihályi’s seminal work Flow: The Psychology of Optimal Experience.

Humans meet Machines

We’ve been talking about learning as if it only happens with humans, but with the rise of silicon cognition (AI) we may have to shift our ideas around a little bit. A while ago a deep learning system called “Alpha Go” used just this reinforcement learning approach to beat the best Go player in the world. The system not only played against younger versions of itself but also played against the top players in the world to practice. First it played with pretty good players, then the best players in Europe and so on. Working its way farther out on the mastery curve.

There are provocative hints at what could happen when both kinds of learners (humans and machines) get together. For instance, there is evidence that the machine learning system pushed the human virtuosos to get even better. An article in Wired Magazine noted that the 633rd player in the world, once beaten by AlphaGo, had disturbing thoughts and nightmares. What defined his life’s passion was taken away from him with lightning speed. Imagine spending a good part of your life becoming good at something, breaking into the ranks of the top 1,000 performers in the world. Then, it’s just gone. He had a rough time to say the least.

Then he started thinking about the moves the machine learning system played. He realized there was a beauty to the strategies. He started thinking in new ways as he tried to understand an entire landscape of new ideas. He decided to do something curious. He flipped over from the virtuoso side of the curve to the amateur side and joined the Google team to continue playing AlphaGo. By the time the number one player in the world was finally beaten, our intrepid human learner had moved up the ranks from number 633 to being in the top 300 in the world.

The loop between making and learning

Our ability and passion to make things that extend our reach has been a cornerstone of what it means to be human. As we make we learn new things about the world and our place in it. Making and learning are intimately tied together at the intersection of the physical world and the world of ideas. So how will the future of making things change the future of learning things?

New manufacturing technologies that take advantage of robotics and automation like additive and subtractive fabrication tools, and synthetic biology systems are enabling us to shape the world like never before. One of the most foundational new methods for designing and making things is at the intersection of advanced manufacturing, the IoT, and machine learning. It’s called Generative Design.

As we shift the way we make towards collaborative connected services and we “sense” more and more of the world, we are discovering new ways to accelerate learning, both for ourselves as well as our machine learning counterparts.

Here are three examples that hint at something profound on the horizon.

1. Hackrod

In a provocative experiment that my team ran a few years ago, we explored the idea that a product could participate in its own redesign. Along the way the team also learned new ways of thinking about design, not only for the vehicle but for the surrounding ecosystem–from new supply chains to new accessories and environments. They were gaining a sort of “system sense” and exploring far more possibilities per unit time than ever before. While the team had always understood the trade-offs they were making as they played out scenarios, with generative design they were able to get an innate sense of how one change over here rippled through the system over there. They found that the way they made things shifted and the act of defining their goals and “playing” with the system–setting weights on this or that goal higher or lower than some other goals–exposed their tacit assumptions and created a rich feedback loop of discovery. Design became more like Mendel’s exploration of pea pod hybrids. Tweak a characteristic here or there and see how successive generations play out in a game of genetic recombination.

At the beginning of the project Felix Holst, the Chief Product Officer for Hackrod, had concerns that generative design systems would marginalize the role of human creativity in the process. “I thought, that’s it, my career is over.” He noted, “it was so profoundly different than the way I learned to design.” As he explored the complex feedback loop between setting goals, sensing forces, and exploring design possibilities he realized, “the power of generative design, running with cloud processes, outstrips anything a team of human minds alone can come up with.” Soon he began to play with the possibilities and fast became a convert. He now feels like he’s in a renaissance in his career and sees the power that these tools bring to extend his own creativity to new heights. “I can’t look back, the old way of designing is too limiting for me.” What he discovered was that the generative design system was a new kind of team member and that together they could collaborate and go farther and faster than humans (or machines) alone could ever hope to do.

While Hollywood would like us to believe that most innovation comes from the lone mad genius inventor, Felix knew that the challenges ahead required deeper and deeper forms of collaboration. As complexity increases and more and more problems are at the intersections of different disciplines rather than squarely in one or the other, teams will become the central currency of innovation. Those teams may be super powered by fostering a mixed dialog between both human and machine learners.

2. Smart Makerspaces

What if the physical learning spaces that our next generation worked within were aware of the context of a given challenge? What if we thought of the physical environment as a sort of machine for learning as well? If a classroom or maker lab was a “robotic” team mate that sensed and responded to the flow of students and activities–tuning itself to accelerate mastery–and formed predictions in real-time about where the student should go next?

A research project at Autodesk Research gives some tantalizing clues to what such an environment might provide. The research team took a standard Instructable tutorial for building a physical product (with some digital circuitry) and measured how quickly and effectively an amateur could build the project and how many times the maker needed expert intervention.

Then they tested amateurs with an “internet of things enabled” workbench. They put LEDs on each part bin to signal which part or tool to grab, they put sensors on the safety goggles, on the drill, on the soldering iron. Now the environment could detect what parts were on the table or if the participants were wearing protective goggles. The system could also light up a light or flash a warning to guide the amateur along the way.

They ran the trials again and had the system show the right information when they detected the right tool–it flagged that the participant needed to wear goggles when they picked up the drill for instance–and provided just the right video clips or instructional steps at just the right time.

What the researchers discovered was that when the room was a learning environment tuning itself in context to actions–when it effectively “paid attention” to the activities of the student–there was a 47% reduction in expert interventions. The environment helped the amateur move towards mastery in half the normal time. 

3. Seventy is the new twenty

Adam Gazzaley had a problem. As a cognitive neuro-scientist he was often asked to speak about how the brain worked and how the brain’s abilities seem to shift as we age. He looked at distraction and working memory and how the two aspects of the brain intersect. He also knew that as we aged these two networks of neurons lost some of their power. If you’re like me you’ve probably noticed something called “silver moments” where you just can’t recall that notion that was on the tip of your tongue. When Adam was asked to speak about this at conferences hosted by the likes of AARP he always felt like all he could do was share bad news. At your peak, around 20 years old, you had amazing powers of attention and a very strong working memory. But by the time you reached 70 years old, things didn’t look so good. There was a marked decline in both abilities. It isn’t particularly fun being the bearer of bad news. After doing ground-breaking work on cognitive decline he wondered if there was anything that could be done about it.

He and his team devised an experiment that combined everything from classic neurophysiological tools to the latest imaging, sensing, and stimulation methods, like fMRIEEG and TMS. They were interested in not only exploring how our ability to multitask declined with age but also if an adaptive set of game mechanics could train the aging brain and reverse the inevitable decline. The results were stunning to say the least (PDF). Adam could demonstrate that seventy could be the new twenty, that the elder brain–once brought into a zone of desirable difficulty–could be trained in a way that fostered the growth of new neurons and restored the performance capabilities back to the peak of a youthful brain.

Meet the mind of a team

We won’t solve the big challenges of tomorrow by going it alone. A high performing and diverse team will be the principle unit of success. Curiously enough building a high performance team is like creating a whole new mind of its own. How could game-changing work in neuroscience be applied to the birth, nurturing and evolution of a high-performance team?

Herb Simon, the Nobel winning economist and pioneer in complex systems once pointed out that a mind is like a pair of scissors. One blade is the brain running the algorithms (think ants looking for sugar), the other blade is the environment the brain pushes against (think gravity, and physics, and the dirt the ants move within). The richness of an ant hive that creates cities, fights wars, and farms fungus can’t be found in either blade of the scissor but in the emergence of something greater as both blades–the brain and the environment–push against each other. Just as one blade doesn’t make a pair of scissors, a mind can’t be seen without looking at both the brain and environment in concert. Our identity is not all inside our head, we are–in some deeply seated way–ourselves plus the world around us.

K through Gray (Grave?)

Think about the way education worked in the 20th century. We built a supply chain from K through 12 and maybe into a post-secondary degree and then we shipped workers off to help build the industrial revolution. It was structured on an “if-this, then-that” model of our understanding of competencies and assumed that there were a few fixed pathways based on majors and disciplines that were required to reach a given goal (if that goal focused on having the same career for 40+ years.) While it did prepare the way for the road ahead, it may not get us any farther than we’ve already come. The entire system had poor summative assessment models (teach to the test), poor targeting, little personalization and worst of all was open loop. Which means we couldn’t create robust formative feedback cycles to drive improvements in outcomes or react fast enough to the pace of change that Moore(computation) and Metcalfe (connectivity) have brought into the world.

But to address the challenges of today and tomorrow–like the massive resource disruptions that automation is creating in the world–we need to think differently. We need to foster life-long learning from K thru Grave. This is no longer an if/then world, learning comes through a closed-loop social network of skillsets, mindsets, teamsets, and toolsets. It’s not majors, its missions. And to prepare for all the missions, for all the possible future “yous” we need a way to simulate those future “yous,” the infinite yous that might be. I wonder what we could learn from the work going on at Stanford, captured in Bill Burnett’s new book, “Designing Your Life” or the work Ayse Birsel has pioneered around “Designing the Life you Love?”

We’ll need a way to predict and explore and prototype possible future pathways. The bad news is that we’ve got a world-wide system of education that continues to pump out students that are perfectly prepared for a world that won’t exist in five years. The good news is that machine learning makes prediction dirt cheap and the shift of so much learning to the cloud has given us a way to instrument the journey like never before.

Next Chapter? Applying generative design to learning itself. Or what happens when you build a learning engine to take us to the stars?


Leave a Reply

%d bloggers like this: