AI: The Technological Trickster
By ai-depot | October 26, 2002
The Memory or the Storehouse
Some of the same problems we find with perceptions also occur when we look at memory as we define it in the brain and in the computer. While the process seems different between memory and perception, (ie: the former seems to take place entirely within the brain and in the computer), in terms of brain function they may be equivalent except in function. The difference, if any, may be something like that between a storehouse and a library.
Campbell and James Jenkins, among others, agree with William James that “…memory is no different from perception, imagination, comparison or reasoning, apart from the fact that… we attribute [it]… to the past.” (Campbell, p. 222) In the brain this may be true, but it may not be true in the computer where past and present may have no real counterpart.
Marvin Minsky, the inventor of frames and arguably the most advanced researcher in AI, suggests we have in computers a library that begins by reading each book to find a particular piece of information, such as Napoleon’s mother’s name. At the speed of electricity this may not take very long, but is still quite time consuming. His work with frames allows a sort of cross-indexing that makes particular information much easier to find, as well as related materials that may be helpful. But the ultimate would be to be able to walk into a library and shout, Who is Napoleon’s mother?” and have the book with that information jump off the shelf and shout back “Myrtle” or whatever. This is apparently how the brain works, and is at least many orders of magnitude beyond what computers can do at present. Whether a qualitative or quantitative difference, Minsky considers it a goal worth pursuing.
While the computer stores information in the random access memory quite efficiently, and retrieves it at the speed of electricity, this seems to be very different from what the brain does. Ernest Kent notes the brain memory apparently acts as a “…storage of average representations of prototypes…as opposed to storage of unique information…” (Kent, p. 233) Somehow we are able to extract unique information from the generalities that result from axon pathways conditioned by use.
Kent and others have isolated three types of memory, sensory, short-term and long-term, which are roughly equivalent to the recognition, repetition and paraphrasing abilities of the brain. (Kent, p. 237�238, Fischler and Firschein, p. 36) Computers may have only one type of memory that functions, or may be made to function, for all three activities, although it’s not clear at present if the first and third will ever be accurately done by machine. But the human brain certainly isn’t completely accurate in any of these categories, as witnesses in criminal trials make us aware. It’s our general-purpose but fallible memory that allows us to function in a dynamic world.
At present there is little evidence that memory can be localized in a specific site in the brain. It seems to take place over the entire cortex and may depend more on a state of the entire brain than on any one area. Charles Furst has suggested that “…a single memory is stored equipotentially in cortical tissue” (Furst, p. 169) while Isreal Rosenfield feels “Memory is not an exact repetition of an image… but a recategorization.” (Rosenfield, p. 192) Both or neither may be found correct with further study, but the latter view, although it seems to contradict common sense (ie: we seem to remember scenes and events quite accurately, may even picture them in our heads), is likely the most realistic interpretation of what happens when we remember. But as Decartes noted, our mind plays tricks on us constantly.
Rosenfield’s general hypothesis is that localization of function is not how the brain works. Thus, “…there is no recollection without content…[which] must, of necessity, constantly change…[which implies] there can never be a fixed… memory.” (Rosenfield, p. 80) How this ability could be duplicated in computer memory is not clear. Some researchers are attempting to program very large computers with a variety of experiential knowledge, a large worldview that may offer a sufficiently large database for this type of activity. But they have been about ten years from their goal for over twenty years now.
If Rosenfield is correct, we may assume “There are no specific recollections in our brains, there are only the means of reorganizing past impressions…” (Rosenfield, p. 76) While a computer memory may be able to manipulate accurate data with little or no error compared to a human memory, it is unclear how the computer may be made to evaluate and manipulate past impressions so as to retrieve a precise picture or impression of a past activity and insert it into a present context. Evolved brains may be rather poor at this, but they do it every day with no apparent effort.
Jenkins discovered in his research that “…remembering takes place in a psychological and physical context which not only varies in itself but is embedded in a wider and wider set of contexts…” (Campbell, p. 220) And the contexts are constantly changing, which requires feedback loops and redundancies that constantly check and correct our perceptions and our memories in a flow of electrochemical activity. This may seem to result in inaccuracies in our ability to perceive and remember the world, but it is probably several orders of magnitude better than the most advanced computer in interpreting a changing world.
Campbell suggests that “In remembering, the brain …cuts down the value of W in Shannon’s equation S = K log W and therefore reduces the entropy. If the items display no obvious relationships… the brain will invent relationships…” The equation also suggests “…unless uncertainty exists first, there can be no information.” (Campbell, p. 215-216, 254) Shannon’s equation aside, the ability to invent relationships, which may be necessary for the operation of memory without degenerating into concrete in a changing and uncertain world, seems beyond the abilities of present machines.
Tags: none
Category: essay |