markh@csd4.csd.uwm.edu (Mark William Hopkins) (12/22/90)
In article <37111@cup.portal.com> mmm@cup.portal.com (Mark Robert Thorson) writes: (Bricklayers who remember details of the brick they've laid years back): >This is not at all the same thing as having a photographic memory >of the face of each brick. The bricklayer is remembering details >he used to make his decisions, so he is remembering the bricks at the >resolution at which for example a master chessplayer would >remember a chess position, rather than the detail of an actual mental >photograph (if such a thing could exist). (A person whose recall years down the line of a document was so perfect, he could identify the document he was presented with as being a forgery due to a minute difference): >The other phenomenon described by McCulloch is the more >well-known example of a person who claims to have photographic >memory. I believe this is nothing more than a person who has >developed a mental exercise which happens to be useful for >providing memory cues... Oh no, I actually think it IS photographic memory in the sense that the image perceived at the time is actually being faithfully reproduced. You see, in the process of perception itself, the brain actually carries out a kind of image compression which also has the by-product of extracting features (hence the observations you note), among other things. During reproduction time, the compressed image is decompressed, hence the total recall. You can best model the process by a two-stage neural net. The first stage performs the compression. Thus its input consists of an enormous number of nodes, and output a significantly smaller number. The second stage performs the decommpression. It takes the output of the first stage and tries to reproduce, as its output, the original input. The by-product (feature extraction) is sitting right there between the first and second stages. The Great Designer saw this unexpected extra and said "It's not a bug, it's a feature!" In the early stages of development, this neural net can be trained (say by backpropagation) so as to minimize the difference between first stage input and second stage output, thus training the identity function on it. If backpropagation is used in modelling this phenomenon, it will be one of the few instances where backpropagation can be used for unsupervised learning. The significance of being able to emulate the identity function this way is the bottleneck that has to be passed through between the first and second stages. Spontaneous feature discovery/extraction is forced on the neural net. What you will find in the brain, that corresponds to this, is RECURRENCE along the visual pathway. That is, connections that go toward the visual cortex from the eye, AND connections that go from the visual cortex in the direction of the eye. I bet you that the recurrence is there to perform the function described above.