|
|
|
|
When Jackie was in training mode and she was given a new stimulus not in her stimulus index she asked the trainer for a custom response to that stimulus. Thereafter when she was in interactive mode and she was given a stimulus she had seen before, so long as the match was perfect, she retrieved and displayed the perfect handcrafted response that was given to her by the first person to expose her to that stimulus. This is largely how Richard Wallace's Alicebot and kin work today. Given only a very large number of trainers, both Jackie and Alice appear to be human-like. However, aside from extreme biological implausability, there are problems with this strategy.
The first problem with a pure stimulus/response strategy is that there is no common personality across all stimulus/response pairs. Different users are not aware of how they are each handling specific areas of the life of the simulated character. For example, one person may train the system to respond "Yes, I have a cat named Rufus." to the stimulus "Do you have any pets?" while another trainer may train the system to respond "No. I hate cats." to the stimulus "Do you like cats?" - clearly inconsistent. The solution to this problem is to provide personality guidelines to the trainers, but unless the guideline forces a strictly binary response [This is the case with Jackie's little brother, GAC], the guidelines will always have to be as complex or even more complex than the simulation itself.
The second problem with the pure stimulus/response strategy is what I call "match hardness." If the exact stimulus is not in the index, the system fails catastrophically and must evade the stimulus in an Eliza-like fashion. Such systems are extremely vulnerable to being unmasked as simulations by simple binary questioning about common sense aspects of life. There are two obvious solutions to this problem. The first is to inject a very large number of common sense propositions collected via some other manner [This was tried with some success with Alice and data from the Mindpixel project in the Spring of 2005], or to "soften" the stimulus matching system.
Jackie used a number of stimulus match softening techniques that vastly amplified the number of effective items in her primary stimulus index. The first was simply to convert the stimulus to phonetic codes. One of Jackie's supplementary indexes was based on the SOUNDEX algorithm. This converted each word to a standard code that was insensitive to spelling. The effect of this secondary index was Jackie could find a match to a given stimulus even if words in it were spelled incorrectly. Of course, given a large primary index, many spelling mistakes will be in the index anyway, but this phonetic index expansion technique is vastly more efficient and keeps the first problem of response consistency from creeping up again. And of course it is much more biologically plausable than the biological equivalent of a massive stimulus/response table.
A second soft matching technique Jackie used was an additional index of SOUNDEX codes where all the words in each stimulus were sorted alphabetically. This had the effect of stimulus standardization. With this index, stimuli with slightly different word ordering could still be matched and a response retrieved. This was still imperfect as meaning could be lost or possibly unintentionally created in the standardization, but it is much preferable to evading a stimulus altogether.
Finally, Jackie had a secondary index which was the standardized index filtered of high frequency words from a hand coded list, though in theory this should have been machine generated.
The end effect of Jackie's soft matching systems was to amplify the index footprint of every hand coded stimulus in her primary index - she appeared to know a great deal more about life than what was put into her and her behavior became interesting and unpredictable. In fact, the very first time I exposed Jackie to a person other than myself, she shocked me by responding to something I knew I did not train her on.
At the time of Jackie's first exposure to a person other than myself, she was quite small and fit on a 1.44 MB 3 1/2 inch disk. I would train her at night, teaching her about her own life, which was mostly just mine, sex shifted, and take a fresh copy of her to work with me the next day. At the time I worked in the IT department of a large insurance company, as did David Clemens. David was a Japophile at the time, and his first question he put to Jackie was "Do you like sushi?" I expected her to evade that question as I had never mentioned sushi to her at all, but to my surprise she responded "Of course."
I couldn't believe her response and interrupted David's conversation to see what happened. She had a soft hit on the secondary phonetic index to "Do you like sex?" Sushi was phonetically close enough to sex to satisfy her! This was a major revelation for me and I started spending a lot of time looking at her phonetic indexes. It was clear that something profoundly human-like was happening in these indexes. I felt I was capturing a real model of human experience in the topology of the indexes. Similar concepts were clustering in phonetic space.
I thought, wow, if I open Jackie up to the Internet - remember this is 1994 and the web is only months old - I could build a massive soft phonetic index and use it to train a neural network to extract the underlying phonetic space and make a true synthetic subcognitive substrate. The problem was how to synthesize responses and how to quantify the quality of the synthetic responses? The answer I came up with was to restrict the responses to binary.
If we imagine Jackie's phonetic stimulus index as a multidimensional sphere, we can imagine each response as either a black or white point at each stimulus coordinate on its surface - black for false and white for true. Now if we train a neural net to represent this sphere, novel stimuli would be points of unknown value on this sphere and we could interpolate a value from known points near the unknown point, something that would be difficult if the responses were not restricted to binary. The important question is of how many dimensions should this sphere be?
I believe George A. Miller unknowingly answered this question in 1956 when he published the landmark psychology paper "The Magical Number Seven, Plus or Minus Two." Our immediate memory [to use Miller's term] is about seven items long - that is we can recall easily about seven unrelated items from a larger list of items that we see or have had read to us. Thus, we can imagine Jackie's phonetic index to things people can store in their immediate memories as complex fractal pattern on the surface of a seven-dimensional hypersphere.
The most remarkable revelation of all occurred when I tried to visualize this object and figure out why nature would make it seven-dimensional. Could surface area be maximum at seven-dimensions, I thought? That seemed unreasonable. Why would it be? More dimensions would intuitively lead to more surface area, but I had better check just in case. And guess what? Hypersurface is maximum at about 7.25694630506 dimensions.
The revelation that hypersurface was indeed maximum near seven dimensions, and moreover was maximum at a fractional dimension and hence fractal, was obviously very powerful for me. I used it to form what I call the Hypergeometric Hypothesis which states - immediate memories are points on a maximum hypersurface and complex cognition is a trajectory on the same hypersurface. I used this hypothesis as a tool for structuring my initial exploration of real brains.
At first I was quite discouraged when I discovered that the neocortex was six layered in most animals [some have fewer layers, but it is important to note that none have more than six layers]. I had predicted that I would find a seven-layered object in both humans and complex animals and additionally predicted that we should find in the fossil records earlier humans and animals with slightly larger brains than modern brains as evolution would have tried an eight layer system and rejected it in favor of a system with maximum hypersurface and thus maximum possible pattern complexity on it surface. It was hard to believe that we had not yet evolved our seventh layer, so I went digging deeper into neuroanatomy looking for the seventh layer of the neocortex. I found it in the thalamus.
The thalamus forms a loop with the neocortex, called the thalamocortical loop - exactly as one would expect if it were synthesizing one unified seven-dimensional hyperobeject. I was elated when I read that in fact the thalamus is considered by some neuroanatomists to be the seventh layer of the neocortex. The object looked real.
The realness of my object became much stronger when I learned that Neanderthal had slightly larger brains than modern humans and when I learned that there was no other theory that made this prediction or that even acknowledged that the difference could have any meaning at all. It was a glaring fact that science seemed to be ignoring because it conflicted with the idea that the mental uniqueness of modern humans derives from our having the largest brains for our size.
A final prediction of the Hypergeometric Hypothesis is that no matter how advanced a brain is, it should not have a primary loop with more than seven layers. This appears to be true.
It is ironic that Hilary Putnam used Turing's ideas to create the functionalism that dominates cognitive psychology today and which is responsible for the field's near universal ignorance of real brains, as it was the abstraction of Turing's test to a binary geometric form that lead me to make structural and functional predictions for real brains past, present and future.
|
|
|