Most machines that people build stop working when their parts break down. Isn't it amazing that our minds can keep on functioning while they're making changes in themselves? Indeed, they must, since minds can't simply shut down work when closed for renovations. But how do we keep functioning while vital parts are modified — or even lost? It's a fact that our brains can keep on working well in spite of injuries in which great multitudes of cells are killed. How could anything be so robust? Here are some possibilities:
Duplication. It is possible to design a machine so that every one of its functions is embodied in several duplicated agents, in different places. Then, if any agent is disabled, one of its duplicates can be made to take over. A machine based on this duplication-scheme could be surprisingly robust. For example, suppose that every function were duplicated in ten agents. If an accident were to destroy half the agents of that machine, the chance that any particular function would be entirely lost is the same as the chance that ten tossed coins would all come up tails — that is, less than one chance in a thousand. And many regions of the human brain do indeed have several duplicates.
Self-Repair. Many of the body's organs can regenerate — that is, they can replace whichever parts are lost to injury or disease. However, brain cells do not usually share this ability. Consequently, healing cannot be the basis of much of the brain's robustness. This makes one wonder why an organ as vital as the brain has evolved to be less able than other organs to repair or replace its broken parts. Presumably, this is because it simply wouldn't help to replace individual brain-agents — unless the same healing process could also restore all the learned connections among those agents. Since it is those networks that embody what we've learned, merely to replace their separate parts would not restore the functions that were lost.
Distributed Processes. It is possible to build machines in which no function is located in any one specific place. Instead, each function is spread out over a range of locations, so that each part's activity contributes a little to each of several different functions. Then the destruction of any small portion will not destroy any function entirely but will only cause small impairments to many different functions.
Accumulation. I'm sure that all of the above methods are employed in our brains. But we also have another source of robustness that offers more advantages. Consider any learning-scheme that begins by using the method of accumulation — in which each agent tends to accumulate a family of subagents that can accomplish that agent's goals in several ways. Later, if any of those subagents become impaired, their supervisor will still be able to accomplish its job, because other of its subagents will remain to do that job, albeit in different ways. So accumulation — the very simplest kind of learning — provides both robustness and versatility. Our learning- systems can build up centers of diversity in which each agent is equipped with various alternatives. When such a center is damaged, the effects may scarcely begin to show until the system's reserves are nearly exhausted.