One can divide artificial general intelligence researchers into reverse and forward engineers. The reverse engineers believe that the fastest route to AGI is to dissect, study and understand mother-nature's creation. The forward engineers believe that we can devise new algorithms running on silicon to come up with a solution.

For most of my years thinking about AGI I believed in the forward engineering approach. Yet, in "The Singularity Is Near" Ray Kurtzweil presents a somewhat convincing view that reverse engineering is the way to go.

Let's look at some other bio-inspired engineering feats:

Natural Device	Artificial Device	Engineering Type
Horse	Automobile	Forward
Bird	Airplane	Forward
Log/Rock	Table / Chair	Forward
Sun	Light bulb	Forward
Fire	Stove / Oven	Transfer / Forward
Ice	Refrigerator	Forward
Cuped hands	Glass	Forward
Stream	Plumbing	Forward
Cave	Home	Forward

In the above there are no cases of reverse engineering. In the case of fire versus the stove / oven we can consider a special case of "transfer" engineering whereby the early artificial devices used the unmodified bio process as in a wood burning stove. But transfer engineering for AGI does not seem viable to me.

It is certainly arguable that the brain is unique with respect to the above listed mechanisms and so such an argument is suspect. However the extra complexity of the brain is equally a handicap on both the reverse and forward engineering sides, suggesting that in the end forward engineering may again win the race.

The two approaches each have certain benefits:

Reverse engineering has the benefit of relatively steady progress that will reach the destination is a relatively controlled predictable way because we will learn a steady stream of information the more we study the brain and all work is focused on one approach. On the downside, reverse engineering may require much more work than is really necessary. For example, it would have taken much longer to get to the first flying devices if the Wright Brothers had to understand everything about the inner workings of a bird. And even after understanding everything about a bird we would then have the task of "optimizing" the design for higher performance and in the end create a forward engineered plane or jet made of metal, engines and props. Furthermore, not doubt there is a lot (if not the vast majority) of baggage associated with the demands of living in the real world - hard code for numerous functions that cannot wait for a learning process and for optimization.

Forward engineering has the disadvantage of uncertainty as to whether humans have the mental facilities to make the needed algorithmic leaps. Yet the endless papers generated by researchers also suggest an inevitable progress to the goal. And forward engineering has the advantage of not being limited by the severe restrictions of evolution-engineering.

As a community, of course a hybrid approach is underway. Forward engineers take inspiration and ideas from the reverse engineers and to a lesser extent the reverse engineers develop models based on forward-engineered disciplines such as math and logic circuits. As a community the hybrid approach is destined to continue.

As an individual researcher the choice remains open. 

Posted 9/15/07