An instructive system in this area is Eric Mueller's DAYDREAMER. https://www.facebook.com/AttentionManagementArchitectures/photos/a.217364258285578.55073.203359906352680/254842044537799/?type=3&theater You may also want to check out this paper I wrote a couple years back, it may be helpful. http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6245681&url=http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6245681 The Mental Simulation / Daydreaming pattern is attached for your reference. For my part, I think you're talking about perception, replay, and mental simulation. Let's use a neural network (or two) as your current model. And the same neural network, or another one as your forward model. Let's call running the neural network "backwards" (upwards) perception, and running it forwards (downwards) replay (I'm not sure what the real technical terms are). Let's assume also that we have a constructive network rather than a fixed network.this means that nodes in the network are created, and the network is constructed rather than being wired a priori. This means, as perceptions are fed in, the networkorganizes itself around temporal and spatial associations, building hierarchical (and heterarchical) nodes that represent sequential and concurrent activation. If this network is continuously fed with input images it should create new nodes at varying levels in the heterarchy. At some point we should be able to replay starting from higher nodes in the hierarchy and projecting or activating downward to lower nodes. This should generate imagesthat can be interpreted (seen imaginatively) by the cognitive system. You can createa system where nodes in the upper levels, if stimulated, will activate the lower levelnodes and replay image fragments of the originally perceived frames. Finally, mental simulation would be another piece of the puzzle. Mental simulation involves creating a forward model based on current input. You perceive a frame of data, store that frame, take a mental action, for example a rotation or some othertransformation of the frame, and store the result of the mental action as the expectedforward model. You can continue to take these mental actions / transformations and eventually you have a movie. The key is taking the mental actions and inhibiting any real motor action to the executor (see the attached image).
Some simulation architectures are here: Demiris and Johnson https://www.facebook.com/AttentionManagementArchitectures/photos/a.217364258285578.55073.203359906352680/226004270754910/?type=3&theater Murray Shanahan: https://www.facebook.com/AttentionManagementArchitectures/photos/a.217364258285578.55073.203359906352680/226004370754900/?type=3&theater https://www.facebook.com/AttentionManagementArchitectures/photos/a.217364258285578.55073.203359906352680/226004407421563/?type=3&theater (combined) https://www.facebook.com/AttentionManagementArchitectures/photos/a.217364258285578.55073.203359906352680/226004387421565/?type=3&theater Some other systems can be found here: https://www.facebook.com/media/set/?set=a.217364258285578.55073.203359906352680&type=3 But essentially you'd want to be able to integrate perceiving, with mental simulation and replay. It could be in that order, where you do some perception on a neural network (the model) and then take some mental actions (transformations) on the perceived images, store them in the forward model, and replay at the same time or a later time. I think the essence of imagination is being able to replay or step through the forward model,activating those nodes in the network (replaying), for some number of cycles governed by somesimulation supervisor. At least that's how it works in my system. (PAM.P2). I'm sure a real, full blown imagination system will be much more complicated than the pattern depicted above, but you've gotta start somewhere. Let me know how this contrasts with what you were thinking. Cheers, ~PM ---------------- > Date: Wed, 17 Sep 2014 22:19:50 -0400 > From: [email protected] > To: [email protected] > CC: [email protected] > Subject: [agi] Reverse engineering the imagination. > > om > > The most pressing issue in my own AGI research, as lackadaisical and > informal as it is, is how the imagination (which is the same thing as > perception) actually works. Solving this might just bring down the walls > of Jericho here. Oops! I just said something really important as an aside. > > Imagination == perception. !!!!!! > > Perception == imagination. !!!!!! > > > Any AGI system that claims to have one but not the other is almost > certainly a fraud. =| > > > The most practical of the questions one could ask about this is how it > works. In the most general sense, the imagination creates the perception > of either real or, well, imagined subjects. We know that it works well > enough for vision that it can prime the visual systems for the apearance > of the actual stimuli that, on short timescales, tiny variations from > the expected input can be detected immediately. Yet, even still this > appears to be far beyond the capabilities of any conventional approach > to computer graphics. Conventional rasterization does not seem to > capture the subtlties of the world and is much too "hacky" a solution. > Raytracing is much more elegant but is not neurologically plausible. > > What we seem to be seeing is a system involving composition of *cough* > patterns that is crudely similar to what you get when you run a neural > network forwards (typically neural network systems are run ass-backwards > because that's more intuitive to ppl with letters after their names). > The most similar system that I am aware of is the fourier transform, > which is where I've picked up my trail of breadcrumbs. This seems to be > a promising approach because there is strong neurological evidence that > the brain performs a frequency domain analysis on hearing, and the same > principle probably applies, to some extent, to other senses too. > > Anyway, Anyone who's not a Kerbal Space Program obsessed dunce like me > can probably run off now and figure out how to tie a FFT-like system > into something that can manipulate abstraction heirarchies and then go > and collect the $10 billion for inventing AGI.... > > Incremental to that goal is the project of developing rendering systems > that work like the imagination does, using frequency composition instead > of rasterization or raytracing... > > -- > IQ is a measure of how stupid you feel. > > Powers are not rights. > > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/19999924-4a978ccc > Modify Your Subscription: https://www.listbox.com/member/?&; > 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-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
