MIMICKING AI’S MISTAKES: WHAT CAN WE LEARN

Artificial intelligence (AI) gets a lot of headlines these days, mostly for its potential applications to everything from making pizza to solving generations-old problems on farms around the world.

 

While the massive data underpinning the technology represents a lot of new solutions, it’s also not without flaws. “Teaching” an AI algorithm sometimes is akin to teaching a 5-year-old how to tie his or her shoelaces. They get the basic concept, but there are a lot of misaligned bunny ears and twisted-up knots on the way to mastery. It’s all a part of how the human mind learns — through practice, failures, more practice and finally fewer failures.

 

Artificial intelligence is no different. Algorithms like Google BigGAN represent amazing potential for the future of a converged digital and analog world. But, just like when we all learned to tie our shoelaces, BigGAN has its screw-ups along the way to what will ultimately be mastery (Janelle Shane is a research scientist who studies neural networks and has a ton of examples here). That doesn’t mean we can’t learn something in those screw-ups that can be applied in new, different ways while en-route to the desired end result (here is a massive collection of images, both realistic and not, from Google’s algorithm).

 

With machine learning like this, it’s all about the logic. And, algorithms and neural networks sometimes use logic that literally can’t be comprehended by the human brain. Sometimes, what’s logical to the machine is indescribable to its human manager. That’s what makes this so fascinating: Neural networks could be discovering things through algorithmic “thought” that human beings never have.

 

And, what is construed as a “mistake” by an algorithm — think an image that the system creates that doesn’t quite match up with reality — could actually be a solution. Biomimicry is the idea of designing materials or machines with nature as the primary design principle. Think a honeycomb structure for building construction that’s exponentially stronger than another similar structure because of the structure invented by bees. Or, how to simulate the vision of a mantis shrimp to vastly improve upon what we can see as human beings.

 

Biomimicry relies on what’s out there in the natural world for its inspiration. But, what if we had reams of neural network mistakes from which to choose? AI-mimicry would draw from the myriad ways neural networks see and perceive things based on some kind of human input, be it a photo, sound, set of instructions or even smell.

 

So, say you want to design a new tractor. Create a dialog with your AI algorithm by introducing content. Say, you show it a photo of a tractor and express what it does, how it does it and what its critical components are. Then, enable the algorithm to use its logic to reproduce that machine. In the thousands upon thousands of images it can return, there are likely to be screw-ups. Versions of the tractor that look just slightly off from reality. Maybe one has a tire attached laterally to the three-point hitch. Maybe it’s missing a windscreen. Maybe it is upside-down.

 

Many of these design misperceptions will be totally flawed and set apart from reality, rendering them completely useless and pointless. But, neural networks have a major advantage. Volume. The calculations required for algorithmic reproduction are possible in massive numbers. Remember the old “Infinite Monkey Theorem?” Just consider a neural network a collection of monkeys on typewriters way, way larger than ever before.

 

The same can be true of a neural network: Give it the right input, time and energy, and it could spit out something that’s both completely absurd and ungrounded from reality, but also yields a new design for a tractor, for example, that’s completely different from any other machine of its kind in the world, yet exponentially more effective than conventional tractors. An AI algorithm may not understand why a tractor has its drive wheels in the back, engine in the front and square windows around the cab. It may, however, understand an entirely different design aesthetic or via a mistake, come up with something that may not even remotely resemble a tractor today, but is way more capable than today’s machines.

 

This means one day, the design of things like building materials and farm machinery may be derived not from what it’s been in the last few generations it’s been around, but on a new logic we human beings can’t even understand. But, that logic could become the basis for a new design aesthetic that has a new level of efficiency, proficiency and ability that we’ve not been able to even fathom.

 

It’s a pretty powerful idea. Harnessing the power of such a system — using “AI-mimicry” as the basis for machinery design — holds enormous potential. The agricultural companies who can both put these types of systems to work and, more importantly, translate their output — whether by design or mistake — will be the ones who lead this change.

Contact Us