Piaget, something like that, yes. I am a little reluctant to commit to a firm opinion until I have a chance to make a computer model that allows me to play with the invreps and episodic memories. My hope is that the model will tell me how they *should* work. I then I will ask somebody to compare the conclusions with actual observation. I follow a constructivist approach: prediction from theory first, then experiment, then comparison.
Anastasios, you are not familiar with my thinking, that's why the confusion. I have a theory based in causality, and I draw conclusions from there. I wish I could tell you "here, read this." I have published several papers, but for now none of them is specific enough to tell you that. Just to mention some differences with my AGI. There is no geometry, addition, multiplication, axes, planes, rotations, etc. All that has to be learned just like you and I learned it in class. Invariants are not "defined", they emerge from experience captured by sensors, for example. There is no program, I do not tell the computer "how to" do things like average colors, or deal with pixels, or recognize anything. It is very different. The computer gets signals from the sensors and infers the invariant representations. They appear as a hierarchy of blocks, and each block has a structure, and so on. Then, only then, I can begin to understand myself and draw meaning. Assign names so you and I can communicate about the blocks, etc. In the traditional approach, your eyes and your brain do all that work and then *tell* the computer how to simulate it by way of a program. In my approach, there is no program (well, there is one, but only to deal with I/O and to minimize the functional). It is very simple, it is an AGI, and some simple things I already have working on my PC. Not 100M pixels, but I can do 1433. Barely enough to make one invrep, barely, but not to actually work with them. The foundations are in theoretical Physics. In Physics, everyone knows that symmetries in a physical system give rise to "conserved quantities." This is may be the most important principle in nature. The conserved quantities depend on the state variables of the system but remain invariant under the dynamics of the system. When I got into AGI and learned about the invreps, it struck me that invreps and conserved quantities are the same thing. And they are, indeed. So now, there is a way to actually *calculate* the invreps, mathematically, from input info obtained from sensors. Now, here is my claim. I believe, but do not claim, that this is the only way to do invreps. But I claim, this time I do claim, that it is very important to build a full-size computational experiment to verify what I say. One with say 1M pixels. For now. Naoya, Yes, to your question, but there is a problem. You, I mean a human, has to tell the computer how to generate those patterns. In the most basic approach, you tell it how to average pixels, how to define invariants, etc. Then, it gets more sophisticated. You, the human, do the patterns or compression in your brain, gain experience about how to do that (you learn, not the AGI), derive rules, and write a program with the rules. The computer then looks as if it were doing the patterns and the compression, but it is still you doing it. Then you can make rules for rules, etc. You know how that works. And you know the limitations. In my approach, I don't tell nothing to the computer. I just flood it with causal relations from sensors, and let it do the inference. Again, this is in theory, my actual little experiments, in my little PC, can not compete with traditional pattern recognition, where a monumental effort was invested for decades. Sergio ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/18883996-f0d58d57 Modify Your Subscription: https://www.listbox.com/member/?&; d2 Powered by Listbox: http://www.listbox.com ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
