AI Taking Baby Steps

Student Media Copy Editor Thomas Breeden

Ben Hirschler reported on Reuters.com on April 2 that Ross King and colleagues at Aberystwyth University in Wales have created a robot that can perform experiments on yeast metabolism and then use the results to design and carry out further experiments.

This marks the first time a robot has made a new scientific discovery unaided by man. The creation of a machine that can use reason to formulate its own theories is a great advance in the field of artificial intelligence.

Hod Lipson, who has a doctorate in mechanical engineering from Technion, is the current director of Cornell University’s Computational Synthesis Lab and is responsible for significant AI advancements of his own. Lipson helped develop a computer program that was able to figure out fundamental laws of physics by observing a swinging double pendulum.

Without any prior instruction in physics, this computer program was able to observe moving bodies and learn everything Newton figured out and then some, all by making and testing calculations. This program may not have figured out anything we didn’t already know, but its accomplishment is still a breakthrough in AI development.

While robot scientists today are only teaching themselves physics and discovering new facts about the genetic make-up of yeast, the next few decades may yield AI systems capable of much greater discovery.

The ability to reason and theorize combined with the ability to process absurdly large bodies of data will allow these robot scientists to map out complex geo/bio/meteorological systems and design cheap, efficient drugs or energy sources. If these sorts of advancements are indeed decades, rather than centuries away, we as a society will need to be aware of what direction AI is headed and what challenges it will pose to how we understand our world and our place in it.

The human body is made up of billions upon billions of biomolecules that perform tasks and interact together in complex, intricate ways. These biomolecules are like tiny machines that together, form the most complicated machine ever, your body.

This is only an analogy, and a misleading one at that, because we are not like machines. Machines are like us. Or at least, they will be, as engineers and programmers begin to base their designs on more complicated biological models.

Rodney Brooks, a professor at MIT, tells us that building a giant computer brain that knows everything we know will do us no good. We need a robot that learns through trial and error, like we do. Robots today are usually designed to do very specific task, like work on an assembly line or dance in Japanese business meetings.

For AI to develop, computers will have to be able to learn and adapt to their environment, using sensory information to understand their surroundings, to be aware of themselves and their world. Scientists today are building robots with odd-shaped bodies and telling them to get from point A to B, without any sort of programming that tells the robot how to move its limbs.

The ability to locomote is the first step toward developing intelligence as we know it and in his essay, “I, Rodney Brooks, Am a Robot,” Brooks notes four distinct marks of intelligence possessed by most human children.

The first is object recognition. Children are able to categorize types of things they may have never seen before but have defined for them. For example, children can categorize different types of shoes as shoes, without having ever seen loafers or cleats before.

Second are language capabilities. Young children have little trouble using complete clauses; they can understand idioms and metaphors, recognize accents and even correct non-native speakers, inferring what was actually meant in a grammatically incorrect utterance.

Third is manual dexterity. Children can manipulate objects they’ve never seen and even objects they can not see at all, such as an object in a bag or a pocket. They can also use objects that are flimsy, like shoelaces, and pick up flat objects on flat surfaces, like cards or coins.

Fourth is the ability to know the difference between what you know and what someone else may know based on what you have observed them observing. This requires a deep level of perception that no AI system has today. These other tasks are either very difficult or entirely impossible for robots to complete.

Brooks is also aware that there may be other aspects of intelligence that we don’t understand or that can’t be easily programmed into a computer. Today, Brooks and his students at MIT study the interactions between real humans and humanoid robots.

While our understanding of our own brain is far from complete, it’s possible that if we figure out fundamental rules of intelligence, we can couple them with what we know about the mechanical function of biomolecules and apply these principles to steel and silicon.

We’re already using these principles to fix artificial limbs and sensory organs to ourselves and while artificial life forms are still a ways off, the initial steps are being taken, and as robots are designed to adapt and learn like organic life forms do, the lines may eventually blur. At some point we’ll be forced to consider how we understand human emotion, intelligence and how we distinguish artificial and authentic life.