In what way do you justify your claim that this is an impending paradigm shift? Many researchers have proposed self-organising systems in the past, such as the Neural Net. This has been around for quite a long time, so it is hardly impending. More importantly a paradigm shift should answer all the questions that a prior paradigm answered plus those which it could not. Managing its own growth or addressing scaling limitations were not the particular questions that top-down designed software was aiming to achieve was it?

Excellent question and good points, Phil. For those who are interested in knowing more about our motivation, I refer you to this paper:

http://xenia.media.mit.edu/~kris/ftp/CustructivistAI-BICA-09-Thorisson-Keynote.pdf

Interesting work, Kees! To address Phil's points in some detail, this is an artificial general intelligence (AGI) project and workshop. AGI involves quite a bit more than pattern recognition as we have to consider all necessary cognitive processes and their interoperation rather than focusing on narrow tasks. So while machine-learning methods based on self-organizing principles are not new, these do not target architectural self-growth of complex, heterogeneous systems containing multiple types of functionality. The paradigm shift I mentioned refers to approaches to building AGI systems, not in all aspects of software development. Top-down approaches were not designed to meet the goals of self-growth or scaling, no - but up until now they were the best methodologies we have had.

Are you referring to self-organizing systems in general, Kees? My personal view on AGI is that all you can assume is that you have some stream of experience which ideally a combination of what is experienced from the environment and what is experienced internally. Some ways of acting (actuators, in an abstract sense as they can be "internal") are also required for affecting this stream. At this low level of experience, pattern recognition is useful to find causal relationships and correlations - this supplies input to build models that can predict future events and prescribe actions to reach certain outcomes (to produce particular patterns in the sensory stream if you will). I don't see any way around bootstrapping AGI systems (i.e. providing them with some manually specified information) but during the course of operation this information can be modified or replaced if "better" information is generated by the system. Aside from this bootstrapping, any assumptions that do not apply to a generic environment like I described should be avoided.

Kees, you state "finding ... themselves, rather ... programmers", this might be a true assumption for today. In a few years, programs will search for solutions that others have found and build predictions. We are closer to this now than ever before. i.e. Watson searches a data base for solutions to questions. Robots can see faces, identify people and objects, speak to us and follow voice commands.

Independent thought in people is always based on what we know. We (humans) search our data base (memory) to find experience. Schools give us a lot of ideas to choose from and tools to use for solutions. Most people do not have "independent thought" based on your definition. To change a light bulb they have been taught to unscrew the bulb, and put back the same power light. Even supper comes with a set of instructions for preparation. Are we not the image of our teachers and instructions? There are very few people like DaVinchie who can truly stand up and claim that they think "outside the box", yet for some reason "outside the box" is the primary goal for robots.

I am content to find, or build, a robot that can follow directions and take care of the basic needs we have with out me telling it every move to make. (Good help is difficult to find) A robot that will see the light is out can just replace or repair it. They are here to be human when we can not be there because of hazardous conditions.

If I give a robot independent thought and self awareness, he will probably go on strike for better wages and longer vacations . Robots are designed to simplify our lives, not to replace humanity.

Kees: Our work may differ in that I am focusing on directed self-growth where the system is initially intended to solve a high level task. If you do not have a clear purpose for such a system, chances are you will not end up with something that is practically useful - although I am certainly not saying that line of research (being close to artificial life) is not interesting. In your example, the agent could learn to operate its external actuators with "motor babbling" (experimenting in random ways, essentially) to observe the effects. If the agent does not have some innate (bootstrapped) knowledge of or attraction to food I don't see any other way to learn what it is without trial-and-error type learning (so more experimentation). This type of learning has a big problem however, it is severely limited in situations involving irreversibility. For example, if there is only a single item of food in the environment and the agent consumes it, any opportunity to discover what else it might have been useful for is gone forever.