Aaron is born
In 1973, Ed Feigenbaum, who ran the prestigious AI laboratory at Stanford University, invited Cohen to spend some time there. That year, in Stanford, Aaron was born. The name, he explains, was chosen
In 1973, Ed Feigenbaum, who ran the prestigious AI laboratory at Stanford University, invited Cohen to spend some time there. That year, in Stanford, Aaron was born. The name, he explains, was chosen simply for alphabetical expediency, At first, I called it Aaron because I assumed that Id be writing a series of programs and I might as well start at the beginning. After about six months, I realised it was my own Hebrew given name and I thought: Uh-oh, someones trying to tell me something. The name stuck. And, he adds, the truth is I havent written a series of programs, it has been the same program that has gone through development.
Up until 1990, Aaron was written in C, then Cohen re-wrote it all in Lisp, the favoured language of AI. The great thing about Lisp, he explains is that its capable of expressing abstractions in a way that C is not. C is fine for numerical this and that, but if you want to talk about colour relationships, or you want to talk about values other than numerical values, Lisp is very good at it.
However, Aaron isnt a graphics program. It doesnt come with plug-ins to give your work the authentic look of Van Gogh or Picasso, it doesnt do transparencies or Bezier curves. It isnt a program you can use. What Aaron does is create spontaneous original drawings and paintings. When Cohen runs the program, it will do a sketch on the screen and then colour it in. Then it will do another and another. Theres no telling what it will choose next (within the limits of what it knows) or how it will compose it. Some of what it does is rather dull, most is appealing, and occasionally its totally compelling.
The driving force behind much of this early work was a search for those cognitive primitives. When I began computing, it started to seem to me that there were, in fact, primitives; that we were all in fact made in the same factory, from the same blueprint more or less... and that when I did such and such, youd probably understand it in a certain way. Over a certain level, youre going to bring your own preoccupations to bear, but up to that, theres no question that a closed form is always going to stand for a solid object... unless its presented deliberately as a flat decorative object. The moment you put two little marks on it, it becomes a face. So there are commonalities that exist, and it was notions about those commonalities that drove the early computing work probably for close to a decade.
Feeding into this was his enduring interest in childrens drawings. He had two sons and a daughter born in England to his first wife. When they divorced, he retained custody and brought the children with him to America. Observing their early drawings, hed noticed a transition from simple scribbles to scribbles bounded by an enclosing line, which led to that idea of the closed form always signifying a solid object. At this time, he also started to explore the powerful imagery of prehistoric cave paintings. The stuff I find most interesting isnt clearly representational, he says. Of course, it can often conjure up representational meaning, but that seemed to him to be imposed by the viewer. We dont really have the faintest idea what the people thought they were doing it for, and I think the anthropologists reading representational intent onto those things is quite mistaken and quite misleading. This all fitted in with his view of art as a meaning generator, rather than a meaning communicator.
At this stage, Aaron was producing abstract drawings of lines and simple shapes. Initially, just on the screen, but later though a mobile turtle dragging a pen over a sheet of paper. Some of these Cohen would take and colour in by hand. By the late 1970s, the shapes had become complex, no longer just scattered blobs. Occasionally, its possible to read meanings into them - a cloud perhaps, or a bird - but theyre clearly not representational. But then, why should they be? After all, Harold Cohen had never been a representational painter.
In 1980, all of this changed. Cohen had been invited to do some consulting at the RAND Corporation, a high-tech think-tank in California. One day, the head of the AI group there asked him if he could get the program to make real drawings... like animals, for example. It was a ridiculous question, given Cohens artistic history, but it touched a nerve. It was at a point when I was trying to figure out what young children do when they draw... Then I thought: if I were to describe a very simple core figure, a straight line with two legs at each end and a neck and have the program draw a line around it, the same way a child draws around a scribble, I wonder whether that would get me anywhere. And it did. The first results looked like African bushmens art. Aaron changed the body shape from a line to a block and they suddenly turned into upper palaeolithic. It was astonishing.
Astonishing, yes, but hardly something to set the art world alight. I wasnt about to go to New York with my animal drawings and show the dealers, but I thought Id discovered a sort of free lunch there, in that the description of the core figure plus the algorithm for doing the embodying outline was much more economical than describing the drawing. So I thought, well, thats something worth knowing. In fact, it was a major discovery and since then all of Aarons drawings have used a schematic core figure made solid by a surrounding line.
As soon as these first representational drawings appeared, they precipitated Cohens plunge into 3D: The program was drawing clearly representational things, but it wasnt drawing them in any ambiance, it would draw single figures or whatever, and I felt the need to put the figures into some sort of ambiance. This forced him to tackle the issue of how to put one object in front of another. Before 1980, it wouldnt deal with overlapping at all. If it found it was drawing something and about to bump into something, it would change its mind and go in a different direction to close the form or whatever it was going to do. So he allowed overlapping, And immediately the things did this... he moves his hands to show the sense of depth, ... a very powerful cue to how you read the thing. It was another big developmental step. Now I realised that overlapping is, in fact, not only a way, but the most powerful way of delineating space.
Aarons current knowledge is of a very limited universe - people, trees, plants, pots, blocks and rocks. Thats a slight annoyance to me, says Cohen, but Ive never been able to persuade myself that either the program, me or the world is going to be better off if it knows about how to draw a telephone... Im much more involved with questions like: Yes, but can you make it hold a bunch of flowers, rather than have a bunch of flowers in a pot? For the moment, the answer to that question is No. Aaron has no way of making objects interact physically because it throws away most of the information about each object before it draws the next one.
This leads to a much bigger question thats currently exercising Cohen: how can he give Aaron the ability to assess its own work and learn from it? He describes it as a hideously difficult problem because it would require a memory of what its done before. But what do you want it to remember? That it did two people and a tree? Whats that going to tell it about what to do next? Oh well, in that case, why not just save the whole bitmap? But what the hell is it going to with a bitmap when it comes back to it? Somewhere within each painting lie the critical factors that characterise this particular picture of two people and a tree and make it significantly different from any other. But, for the moment, finding them seems an insurmountable problem.
As we talked, Aaron was busy nex
This article appeared in the April, 2000 issue of PC Authority.
Thoughts on this article? Add a comment below.
Be the first to comment on this article.