Aaron, I am not trying to pretend that I know something about Autoencoders and I am interested in the additional information that you have provided. However, the very term "unsupervised" (that you see in the literature) is relative. (It has to be, but so far that term has been used to suggest that the algorithms can do something that they haven't really done.)
But your statement that they are the only methods which even attempt to solve this problem and that you either have to know in advance to recognize relevant features, or you have to learn from experience," really suggest that there is a fundamental misunderstanding at the root of our disagreement. I will try to explain my opinion about this. If you have the time and you read more about them please tell us more about autoencoders. Have you tried writing one yourself? (Presuming that you haven't), is there a simple model that you could start out with? On Mon, Dec 30, 2013 at 1:33 PM, Aaron Hosford <[email protected]> wrote: > *PM:* > Perhaps you're right about babies planning. It's a bit of a grey area, to > me. Can we really call a newborn's cry a form of "planning", or is it a > simple stimulus-response pairing, a default behavior used in lieu of a plan > when none is available? But it is not long before babies begin to learn to > control their behaviors, and at least by this point, planning is present. > Regardless, I would argue that in order to plan, we must already recognize > some salient features of the environment, and how they relate to each other > temporally & causally to some extent. This calls for some sort of > observational intelligence on which executive intelligence can rely for > relevant information. It seems to me that the most information-dense > features -- the ones that have the highest predictive value, corresponding > roughly to your option #4 -- are the most relevant to planning. > > When I think about planning in an adaptive cognitive system, I imagine > working in the opposite direction, starting with features of the goal state > and using experience to identify other features that lead to the directly > to them, and then recursively applying the same process to those features > treated themselves as (sub)goals. Over time, backwards chaining of > goal-predicting features should build up a forest of predictive features, > with primary goals as roots of the trees in the forest. These feature trees > could then be used by a planning engine as you describe, to identify the > feature chains that carry us from the currently observed features towards > the roots in this forest. > > Or perhaps a better analogy would be a web of features, linked together by > events or behaviors, in which case planning would correspond to path search > in a directed graph where salient features are vertices and > events/behaviors are edges. It would be the job of an observational > intelligence algorithm, something loosely based on autoencoders, PCA, or > other salient feature extraction algorithms, to construct the graph that is > searched by the executive intelligence's planning engine. This approach has > the advantage over the idea of goal-rooted forests of the previous > paragraph that goal states can be changed without affecting the underlying > feature space graph, because goal features don't get special status. > > There is no escaping the fact that at some point, there is a certain > amount of guesswork involved in identifying salient features, because we > can't measure the utility of the features until they have been used by the > planning engine. The only way to do this is either to manually construct an > appropriate feature detector in advance, or use some sort of autoencoder- > or PCA-like algorithm for finding the juiciest morsels of information based > on intrinsic correlations in the sensory and motor data. Once the planning > engine starts utilizing these features, it can start to give feedback to > the system responsible for choosing the most salient ones, and this > information can be used to selectively identify features that specifically > correlate with those already identified as being useful. > > > *Jim:* > Autoencoders, PCA, and their close relatives are the only methods to date > which even *attempt *to solve this problem, to my knowledge. These two > families of algorithms are specifically engineered to the task of > identifying features without supervision. Both are data intensive, and I > think this is probably intrinsic to the task itself. Either you have to > know in advance how to recognize relevant features, or you have to learn > from experience. I think biological organisms have a leg up because we have > evolved specific perceptual mechanisms tailored to the sensory modalities > we utilize. Either we will have to engineer tailored perceptual mechanisms > for the specific sensory modalities of our machines, perhaps by emulating > those of biological brains, or we will have to utilize some sort of > general, unsupervised algorithm comparable to these two. > > The interesting thing about autoencoders, as compared to PCA, is that they > can be stacked to produce higher-level descriptions of the inputs. Each > layer maximizes the descriptive accuracy of its output within the > constraints of limited space, and the smaller the output is constrained to > be, the more general the description becomes. Many on this thread have > already discussed the deep relationship of intelligence to compression. > Autoencoders are simply a way of learning effective > compression/decompression encodings that are optimal for the given data > set. The more compressed the description, the more informative (and > therefore important) each bit of that description becomes. How else could > we possibly identify high level concepts that effectively describe the data > without already knowing something about them in advance? > > > On Mon, Dec 30, 2013 at 10:53 AM, Jim Bromer <[email protected]> wrote: > >> I really do not see how the assumption that an AGI program can detect >> features other than the primitive features of a sensor can be taken as a >> grand stepping off point for the higher level analysis and reasoning that >> you want to design into a program. John Rose talked about the correlative >> structures of the world, and this might be used as a good basis of further >> analysis of how an intermediate stepping off point would work. But the >> idea that something like autoencoders, PCA or perception and categorization >> are sufficient to explain how features may detected from the sensor >> environment should be something that, if the theory was true, would be >> feasible to prove right now - like my claim that an AI program could learn >> to use CF Grammars through trial and error learning is something that >> should be easy to demonstrate. >> >> If the significant features that were needed for an intelligent >> comprehension of events of a data environment were easy to detect then AGI >> would be easy. The whole problem has been that this initial stage of >> the detection of effective features has not been easy. >> >> Jim Bromer >> >> >> On Mon, Dec 30, 2013 at 3:13 AM, Piaget Modeler < >> [email protected]> wrote: >> >>> I agree that goal directed behavior is at the top of the pyramid of >>> functionality, >>> with perception (observation) and categorization (coordination) at the >>> base. >>> >>> I disagree because I do believe Babies plan. I also believe babies >>> featurize their >>> environment. I think they use the features to plan. One basic plan >>> that babies >>> employ to achieve their goals is called "Cry". Babies use that plan in >>> numerous >>> situations, and possibly when all other plans fail. >>> >>> My question was different. It was, given a grand number of features >>> available at any >>> given point, how do we choose the most "relevant" ones in order to >>> formulate the >>> current state. Or in the case of a planner, an initial state. >>> >>> A few approaches exist: (1) Pick a few features at random; (2) sort the >>> features in some >>> way and pick the first N features; (3) pick no features, and instead >>> plan using an empty >>> initial state and a goal state; (4) see if any correlation exists >>> (temporal concurrence, >>> temporal sequence, etc.) between features and the goal, if so, pick the >>> initial state >>> features based on this correlation. Doubtless there are other possible >>> answers >>> as well. >>> >>> Perhaps experimentation with all these approaches should be done to >>> determine which >>> works best. >>> >>> ~PM >>> >>> ------------------------------ >>> Date: Sun, 29 Dec 2013 22:54:58 -0600 >>> >>> Subject: Re: [agi] Problem [Space] formulation >>> From: [email protected] >>> To: [email protected] >>> >>> >>> First comes awareness, then comes play, then comes planning. Babies >>> don't plan anything. They just observe and correlate. They pick out salient >>> features based on what allows them to predict the most about future >>> observations, particularly those relating to reward or suffering. (Human >>> faces and voices, among other things, are probably recognized a priori as >>> salient through hardwired mechanisms.) Play allows the non-goal directed >>> testing of observed correlations between salient features. Planning can >>> only take place after you have started to recognize that salient feature X >>> leads to salient feature Y, which is associated with reward or relief >>> from/avoidance of suffering. Build the foundation first, in the form of a >>> salient feature recognition engine -- an autoencoder* or some type of >>> autocorrelation technique -- and then implement your planning engine on top >>> of that. Goal directed behavior is at the top of a pyramid of functionality >>> with perception and categorization at the base. >>> >>> *Not sure if you're familiar with >>> autoencoders<http://en.wikipedia.org/wiki/Autoencoder>. >>> They are a technique used for deep learning, among other things. The idea >>> is to teach a pair of neural networks or other classification-learning >>> algorithms to learn a compressed, (nearly) lossless encoding, which by >>> necessity pulls out the most salient features of the input, and then to use >>> the compressed representation as the input to the next layer, or in your >>> case, the input to the planning engine. (The decoder is typically thrown >>> away once the encoder/decoder pair has been trained.) It's similar in >>> function to principal component >>> analysis<http://en.wikipedia.org/wiki/Principal_component_analysis>, >>> which could also possibly serve your purposes here. >>> >>> >>> On Sat, Dec 28, 2013 at 11:14 AM, Piaget Modeler < >>> [email protected]> wrote: >>> >>> Thanks Jim, >>> >>> Much to think about... >>> >>> ~PM >>> >>> ------------------------------ >>> Date: Sat, 28 Dec 2013 08:33:22 -0500 >>> Subject: Re: [agi] Problem [Space] formulation >>> From: [email protected] >>> To: [email protected] >>> >>> >>> You could, for example, try to write a planner for solving the problem >>> of selecting the relevant factors to form an initial state (and all the >>> other possible states) for your lower solver to work with. To do this you >>> would probably have to use some kind of abstractions. Since you want it to >>> be based on solid scientific methods, you might try something like, >>> material objects, properties and relations between objects. This you might >>> find would not work so well since you want an AGI program to deal with >>> imaginary things like fantasies and creative solutions to problems. So >>> then you might start by working with something like ideas, concepts and >>> conceptual relations. However, this does not solve the representation >>> problem so then you have to include something about representations. >>> However, you are back to the earlier problem, you cannot 'represent' every >>> thing that you can think about so you might try to find a slightly better >>> word-concept to use, like 'symbols' or 'references'. Even though you >>> cannot 'represent' everything there is about the universe, for example, you >>> can still refer to it. Since you will need a way to represent the symbolic >>> references you might try to rely on grammatical primitives like 'nouns' and >>> 'verbs'. This however implies that every concept-situation you might want >>> your solver to represent could be represented with individual words. This >>> does not quite make sense so you might then use actions and objects as part >>> of your native abstractions for the solver to use in solving the problem of >>> building a state system (or other kind of system) for a higher level (or >>> more primitive) solver to solve. And you might try to think about using >>> something like computational syntax instead of linguistic syntax, since, >>> your solver is going to be a computational program and it will act as if it >>> operating on computational strings anyway. These ideas however, still do >>> not solve the problem because it is much too abstract and too full of >>> gaps. And while the abstractions may seem like they are ideal for >>> representing any possible situation, it turns out they are not because they >>> are themselves possible states of thought. So, for another example, can >>> you have a representation of something that cannot be represented? Well >>> yes, since you can represent a reference to it. But that means that your >>> solver has to be on guard against getting caught into illogical >>> representations that refer to things that cannot be referred to in the way >>> your abstract program is going to try to refer to them. While the >>> comprehension and representation of the entire universe (for example) may >>> not represent much of a problem to anyone other than a fool, it does show >>> how problems of symbolic representations and references can creep into your >>> solver when you start with higher abstractions. So now you need to add an >>> incremental trial and error method to detect such possibilities. And since >>> such things may be hidden by more complicated systems of references the >>> problems cannot be easily detected. And incremental methods may not be >>> strong enough to gain traction for your solver to create possible plan >>> states (or other plan domains) so your solver solver will have to be able >>> to latch onto something that will build traction into its incremental trial >>> and error methods as it progresses in its goal to create a domain that is >>> suitable for a planner. >>> >>> Oops. I am talking too much about my own ideas again. Sorry. However, >>> it all seems so relevant to the situation that you just described, and so >>> relevant to the question of the problem of creating an actual AGI >>> program that I can excuse myself. >>> >>> >>> On Fri, Dec 27, 2013 at 11:34 PM, Piaget Modeler < >>> [email protected]> wrote: >>> >>> So I'm writing my solvers, and I hit the first roadblock... >>> >>> Given (1) a sea of sensory stimuli, (2) a prioritized set of goals, and >>> (3) a possibly empty >>> set of plans, how does one select the relevant stimuli to form an >>> initial state for a problem? >>> >>> Another way to ask the question is how do humans select relevant >>> features from their >>> current environment to be able to formulate or retrieve plans that >>> address their goals? >>> >>> And how many of these environmental features are enough to describe an >>> initial state? >>> >>> So there is PDDL. (Planning Domain Definition Language). But to use >>> PDDL, one has >>> to first solve the problem of describing the relevant features of the >>> environment. >>> >>> How do we come up with these relevant features, to be able to formulate >>> an initial state? >>> >>> Thoughts? >>> >>> ~PM >>> >>> (I forgot what Newell & Simon said about the PSCM & Unified Theories >>> of Cognition. >>> Time for more research...) >>> -- >>> >>> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/23050605-2da819ff> | >> Modify <https://www.listbox.com/member/?&;> Your Subscription >> <http://www.listbox.com> >> > > *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> > <https://www.listbox.com/member/archive/rss/303/24379807-f5817f28> | > Modify<https://www.listbox.com/member/?&;>Your Subscription > <http://www.listbox.com> > -- Jim Bromer ------------------------------------------- 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
