caro thinks

Dream Beginnings , 2002/02/21:15:29

caro's home ~*~ caro's index

Do robots dream of electric sheep? They do if they heard someone talking about them.

artificial intelligence respond responses (thread index) next in thread: michael2002_04_02:21:24 (thread index) next in thread: caro2002_02_22:23:17

New Topic for the journal. Here I'll write notes for my paper on a.i., which I will present at the Second Annual Enlightenment Meeting this June. First, some distinctions need to be made to keep things clear. There are several terms that tend to get confused in discussions of dreams: 'REM sleep', 'brain wave pattern', and 'dream'. REM sleep, or Rapid Eye Movement Sleep, is detected by looking at the sleeping person's eyeballs rolling around under the eyelids. For the purposes of this discussion, this is all that will be meant by REM sleep--that the eyes are visibly rolling around. NREM sleep, or Non-Rapid Eye Movement Sleep, is detected by the absence of rolling of eyeballs under closed eyelids. (In the OCM article on dreams, Iswald says, "There are two kinds of sleep..." Surely much finer distinctions are in order?) Dreams are detected by asking conscious, non-sleeping people who have just awakened (or by their volunteering), generally in verbal form, "what were you dreaming?" or "what thoughts were passing through your mind?" The people then report, verbally, a dream. I sometimes refer to this as the subjective part of the experience, the qualia, if you will, of the neuronal activity. IMPORTANT: People will report dreams when awakened during BOTH kinds of sleep. The difference is that, when awakened during REM sleep, people report more visually-descriptive dreams; when awakened during NREM sleep, they report less visually-descriptive dreams. My Old Theory of Dreams (which I still believe, and now may have a good foundation for, in my New Theory of Intelligence) We don't dream while we are asleep. Upon waking up, we confabulate: i.e., we tell a likely story that makes sense of the neuronal activity that had just been occurring. The origin of this theory is some information about pilots. When pilots stay up in the air for a long time, sleep-deprived and subjected to reduced air pressure, they report that they "saw" gremlins on the instrument panels. Sometimes the gremlins talk to them. Because the reports were so similar, coming from different pilots, the phenomenon was investigated. It was found that the capillaries in their eyeballs burst, causing some optic nerve stimulation but not in a way that the pilots could say they consciously saw. The conclusion was that, in order to make sense of this stimulus, they confabulated a story. Talked to Tom about this last week, who suggested that the reason that they pilots specifically reported seeing gremlins is that the term 'gremlin' has a long history in aviation. Anything that goes wrong in a plane is caused by "gremlins." Thus it would be a word in their ready-to-hand vocabulary, and some disturbance in their visual field could be interpreted as something "going wrong"--and since they were overtired (and possibly partly asleep), it wasn't all that surprising to them to actually start seeing the gremlins that were the source of all the problems. And this inspired me to suggest that, the human brain being set up to distinguish human faces better than pretty much anything, if the pilots had some visual sensation they couldn't immediately resolve in some other way, they might try to find a face in the blur. Gremlins. Completely confabulated in a more or less story-like way, on the basis of neuronal firing that couldn't be resolved. This piece of information suggested that other bizarre phenomena might be similarly caused and explained. Thus, my theory of dreams, which recurred to me recently while I was drawing some diagrams in an attempt to understand and solve what Ted called "the binding problem in artificial intelligence". I'll detour here into my results, which didn't solve the binding problem (although it made me increasingly suspicious that it was a spurious problem akin to Xeno's Paradoxes), but produced a pretty little theory of how to save on processing and storage overhead when developing an artificial intelligence. According to my developing theory of intelligence, we sleep in order to sort through and reorganize what we know. I refer, loosely and metaphorically, to what we know as "the big database." Short-term memory is first in, last out: a stack. Like a can of Pringles Potato Chips, as Mike Young so handily put it when I was explaining stacks to him this weekend. Onto the stack, all day long, data gets pushed in the order in which it occurs: images, nice turns of phrase, people's names, faces--in fact, EVERY thing that we experience, whether we notice it consciously or not. Evidence that short-term memory can be usefully described this way, is this sort of example, which happens to people all the time: Suppose two people are having a long conversation. At some point they stop and say, "Wait a minute! What got us onto this? Weren't we talking about something else? Seem to remember it was important..." In general (and there are exceptions which I will consider separately, because I can use them to help explain the mid-term memory data structure)--in general, people find out how they got here in the conversation by literally retracing their steps, backward, thus: "We started talking about libertarian politics because before that we were talking about Danielle Van Dam and how all of this could have been greatly alleviated if children had chips and transmitters implanted in their bodies and you started getting defensive about surveillance; and we got onto Danielle because I was half-watching the news while we talked, and, uh, before I got distracted by the news we were talking about, uh, Oh! Your ski trip, and the reason we were talking about that was that we were talking about the upcoming meeting and you were wondering whether you had the money to do both... This kind of retracing seems to literally go in order: last in, first out. I'm popping the stack, trying to get to the bottom of the Pringles can so I can see where this conversation began. It is difficult--though not always impossible--to do it in another order. The short-term memory stacks and queues are where we put stuff for temporary storage. These structures are not for understanding, but for brute storage. Later, we will empty them out element by element, and check it against the database for differences and similarities. The big database is a multi-dimensional hash, or, as Ted O'Connor noted this week, it is not really a hash, but more of a soup. My understanding of a soup is that, like a hash, there are keys with associated data, but the structure is multi-directional. This avoids the wasteful problem of having a key refer to data that refers to the same key further down in the multi-dimensional hash. Instead, the data refers back to the key. I don't care how precise this is right now. I'm in a hurry, and there is much to be blurted out. My stacks are really, really full, my processors are working overtime, and I may even be asleep right now. I think I hear them coming to take me away.... Mid-term memory, which may come in several degrees or sizes or structures, is basically a temporary hash separate from the big database (where 'separate' might even mean, 'is one special flagged element of the big hash'). Recently, Ted and Jing Jing were here. Ted started some subject, Jing Jing asked a question about it. Though the subject was mostly unfamiliar to her, she suddenly said, "Oh! This is what Bernard was talking about." and then had something to say about what he said. Where was this stuff? It wasn't in short term memory; it didn't seem she was digging it out by popping a stack. Rather, she listen long enough until Ted said a key word or phrase--and it was a key she had in her mid-term memory hash. The data associated with that key was, in part, Bernard, and what he'd had to say about the topic. Mid-term memory is where we are putting information that we need to access in a flash. Sometimes this is stuff we have already found places for in the big database, but we've pulled back out Mid-term memory is where ALSO we are putting things, sometimes while learning them, that seem important but we don't yet know how to fit them into the big database--or perhaps even where we put things when we know how important they are but we're too tired to make a big, expensive query to the database (and we might say "I don't feel like talking about this (normally favorite) subject right now"). We forego brute efficiency of the stack and take on the higher space overhead of the hash in order to benefit from the random-accessibility of the hash. We can extract the information with lightening speed, even when we don't yet have a full understanding of it in the context of our wider knowledge. Perhaps the rule is even, "If it happened recently, it must be important." or "If it happened recently, then there's more information related to it that is sure to follow." Incidentally, this is another way of putting Price's classic theory of concepts as dispositions, where some concepts are just on the threshhold and ready to be used whereas others are buried more deeply and need more stimulation to be activated. Article Analysis: "Dreaming," by Ian Iswald (professor of psychiatry), OCM, pp 201-203 Let's look at some of the prejudices that are apparent in the article, simply stated without argument as basic fact. The first sentence: "In our sleep we all intermittently experience insanity" (Dreaming, p. 201). This provocative sentence indicates that the author thinks there is no reasonable explanation of the bizarre things that we dream. I think that there is. A couple of paragraphs later, Iswald says "We may need to dream, but no one has yet devised an experiment to see whether we have such psychological needs at night" (Dreaming, p. 202, col. 2). This statement indicates that, because the reported dream is considered a psychological phenomenon, the author's own thought processes are focused on the dream as psychological and important in and of itself, rather than as an effect of something that is important. I think it is an effect of something much more important. Another strange statement regarding data: "While the majority of the REMs certainly cannot be ascribed to scanning the visual field of a dream-world, there are occasional large eye movements that do seem to bear a relation to described dream content" (Dreaming, p. 202, col .2). This statement suggest that Iswald is firmly in the camp that the actual _dream_ occurs during sleep, that the dream itself is the significant, interesting activity, and that the point of the neural activity and rapid eye movement is to cause the dream, rather than to do something important to us. I think that dreams occur because we wake up and, being explainers and shelvers, need to make sense of the activity that had been occurring in our brains. More on this article later. Sleeping, Neuronal Activity, and Dreams Popping the short-term memory stack and checking each element against the whole database is expensive, both in terms of processor overhead and time. Whatever else sleep may be for (healing, rebuilding, growing), this is when we are reexamining our information, finding fits, making connections, and storing. Studies indicate that we remember learned material much better if there is a period of sleep intervening between the time of learning and the time of testing. Why would this be so? If my emerging theory is correct, it is because we need to shut down all systems not directly essential to learn (or surviving), and focus all resources on going over the database, for a long, uninterrupted amount of time. In large part, the bizarreness of dreams provides the inspiration for this emerging theory of intelligence. They give us some insight into how we're doing this thing called knowing. Consider the things you might store in any given day. Consider taking a single thing off the stack of things you stored, and checking it against the whole database. Suppose that I store the fact that Terry wants to go to the nursery to pick out the plants listed on my plan, on Saturday. I see him as he's telling me this, leaning over my patio gate. While sleeping, I get to this piece of data in my short-term memory stack, and check against everything I know. I'm somewhere in the medieval history portion of my database, when the alarm clock rings. I wake up with some disparate piece of data flying around in my head. I try to tell a story to make sense of it. So I say "I was just dreaming that Terry came over to tell me when we were going to the nursery--only I wasn't at my apartment. I was at the philosophy department at Indiana University. I was in Paul Spade's medieval logic class, and suddenly Terry just leaned in the window and started scheduling with me, and for some reason I didn't think this was strange at all...."

Notify me when caro writes again.

Find Enlightenment