I largely agree with you. What you've described is what PAM-P2 is attempting to do.
Stay tuned. > Date: Fri, 29 Jun 2012 09:19:42 -0400 > From: [email protected] > To: [email protected] > Subject: Re: [agi] Building high-level features using large scale > unsupervised learning > > Ben Goertzel wrote: > > How exactly do you suggest to bridge the functionality gap between > > visual pattern recognition and all the other things human beings do? > > =) > > Setting aside problems noted as still being unsolved, here's a crude > sketch of how the system can be organized. For the sake of brevity, only > the cortical-thalamic-cortical system will be considered. > > The first thing to note is that this is an unsupervised pattern learner. > That should be pretty amazing all by itself. The second thing to note is > that all it deals with are vectors of numbers. There is no reason on > earth that it can't be made to work with any conceivable stimulus that > can be encoded as a vector of numbers. There are some serious channel > dependence problems, previously noted, but the basic process is present. > > The third thing to note is that they could run their matrix stack in > reverse and "imagine" what a face looks like. This is critical, > especially for motor control! =P > > This is your basic algorithm. The next challenge is that you need to > break channel dependence and introduce associations between patterns ie > with faces and the various representations of the word "face". I suspect > that once channel dependence is fixed, then, at some high level in the > network, these associations will emerge on their own. > > The next issue is topology. You could organize the topology like the > human brain and, in theory, it should be human equivalent. Motor control > is implemented just like perception. It builds up complex sequences of > actions from simple sequences of actions exactly as complex perceptions > are built up from simple perceptions. To do something, you just run the > stack in reverse, as mentioned above. Combined with channel dependence > and free association, you obtain arbitrary sequences of planned actions. > Actions that are fully learned become habitual (simply initiate the top > level abstraction). Other actions require an iterative system-wide > process for planning, but most of the mechanisms are already present. > > You obtain episodic memory by having a pipeline that associates > concurrent perceptions, which appears to be what the hypocampus does. > > To obtain super-human intelligence, you need to make the topology of the > system adaptive, or even accessible to the system itself. Ideally, you > want a highly redundant, highly distributed, highly parallel and highly > efficient architecture. This architecture does have a second class of > scalability issues, each matrix, at each level of abstraction is of > fixed size, There needs to be a process that simplifies and consolidates > knowledge to a more ideal representation. At that point you're off the > edge of the (metaphorical) napkin I sketched this all out on. =P > > About 80% of everything else you need is already available off the > shelf, the other 20% might have some important, perhaps even difficult, > challenges but then we're talking about emotions and motivation instead > of intelligence. > > -- > E T F > N H E > D E D > > Powers are not rights. > > > > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/19999924-5cfde295 > 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-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
