TRON Legacy
More info i
Credit: Disney

Biodigital jazz, man! Tron: Legacy's ISOs are kind of like real-life biological algorithms

Contributed by
Dec 16, 2020, 3:50 PM EST (Updated)

It's been a decade since we returned to The Grid in Tron: Legacy. The sequel to the classic sci-fi film reunited us with familiar faces, but also introduced some new concepts, all wrapped up in neon and a ridiculously good Daft Punk score.

Among those new ideas was the introduction of ISOs (isomorphic algorithms), artificial entities that emerged naturally within the environment of The Grid. Jeff Bridges' Flynn describes them as "something extraordinary, a miracle." Biodigital jazz, man. Turns out, that idea is based on some very real science.

EVOLUTIONARY ALGORITHMS

So much of our world feels entirely artificial, made up of constructs that don't exist in nature, but that isn't entirely so. In reality, much of our technology takes its inspiration from nature.

Famously, George de Mestral got to work inventing Velcro after noticing the way burrs from burdock plants would get caught on his pants. Natural selection figured out the basics of lightbulbs, sonar, suction cups, and powered flight long before we humans ever got around to tackling them. Biology presents an array of effective blueprints for achieving physical feats, it only stands to reason its processes might be useful in computing, too.

Genetic and evolutionary algorithms borrow from natural selection in order to find solutions to specific problems. Rather than a programmer designing an algorithm from the top down, a population is introduced to a problem and measured against it to determine a "fitness score."

You'll notice as we continue, some familiar language more commonly associated with plant or animal populations.

The fitness score of each member of this mathematical population is calculated by bumping its performance against an intended target. It's expected that, in the beginning, the scores won't be very good. But some will outperform others.

Next, a subset of the population, those with the best fitness scores, are selected and allowed to "reproduce." This can be accomplished in a few ways. Some individuals continue to the next generation unchanged. Others undergo random mutation, where inputs are arbitrarily substituted for new values. Some will be recombined with peers in a process known as "crossover." For instance, half of the inputs from one member of the population are swapped with the corresponding inputs from a second member.

Once this is done, you're left with a second generation and the test begins again. Rinse and repeat. Over time, the fittest solutions are selected for you to move nearer to the desired outcome until finally (hopefully) it is achieved.

These sorts of algorithms can allow for solutions to previously unsolved problems. Alternatively, they may simply present different solutions than we'd previously invented. A straightforward (and hilarious) example involves asking software to create digital animals capable of completing a task, as demonstrated by experiments Karl Sims carried out in the '90s.

Not only does it make for entertaining viewing, it also allows engineers to test machine designs in a virtual environment to see how they perform without needing to build prototypes in the real world.

Credit: Disney

TWO KINDS OF ALGORITHMIC "LIFE"

Research involving evolutionary algorithms has, over recent years, moved closer to what we'd consider life in two important ways: digital and biological.

Earlier this year, Quoc Le, a computer scientist at Google, applied the evolutionary algorithm process outlined above to artificial intelligence. Their process allows them to test thousands of potential algorithms per second until they find one capable of performing a classic AI task, like identifying an image.

This process removes a lot of the manual work usually associated with teaching a neural network to recognize something. To be fair, the solutions they uncovered were simple as compared to the cutting edge of traditional AI, but it showed the process works. The breakthrough is not in the result, but in the process and the proof of concept that under the right pressures, machines are capable of learning on their own.

Le intends to scale up the process, with the hope it can later uncover computational abilities we've not yet solved.

Perhaps more exciting is the work being done with xenobots, biological robots crafted from the cells of African clawed frogs. Scientists studied the elastic properties of clawed frog skin cells and the locomotive potential of heart cells and fed those constraints into an evolutionary algorithm.

Then, using much the same process as was shown in the Karl Sims experiments, researchers allowed their algorithm to determine potential structures capable of moving through a watery environment or even carrying a small payload. The best were selected and the worst weeded out, and, eventually, a group of structures was chosen and constructed using the frogs' cells.

These tiny novel organisms, measuring only a millimeter to a side, were able to move through the environment, manipulate objects around them, and heal when injured. The team believes there may be many real-world applications, including medical therapies and cleaning up pollution like microplastics in the world's oceans.

The practical results of evolutionary algorithms are still in their nascent stages but their potential for changing the way we think about computing, particularly in removing our inherent biases, are tantalizing.