Attention

Over the several decades that I have been doing AI research the biggest problems have been the related issues of control of complexity and focus of attention.  Can my current AI, Asa H, be taught what it should attend to?  Can the most important cases in each level in the memory hierarchy be taught sufficiently high utilities and, most importantly, will this exhibit the required dependence on context? Does having vector utility help? The utility of a given case (pattern/concept) depends upon the context. The problem of control of attention often times didn't arise in simpler domains but is important for operation in the real world.

Examples of vector utility

If you buy a resistor its usefulness in a given electronic circuit depends upon, at least, its electrical resistance, R, and the power it can withstand, P.  It has a vector utility of at least U = (R, P).  The utility might also depend upon the dollar cost of the component and its physical size as well.

If you buy a capacitor its usefulness depends upon, at least, its capacitance, C, and the maximum voltage it can tolerate, V.  Its vector utility is then, at least, U = (C, V). No single scalar utility can be assigned to the capacitor unless you can specify its application.  If the capacitor is to store charge, say as a memory cell, then perhaps a suitable scalar utility would be U = Q = CV.  On the other hand, if the capacitor is intended to store energy, say in a capacitor bank, then perhaps a suitable scalar utility would be U = E = .5 C VV.  If you wish to store charge while using a minimum energy then perhaps U = Q/E = 2/V.  A suitable scalar utility depends upon the context at the moment. A general utility needs to be a vector. (See chapter 2 of my book, Twelve Papers, www.robert-w-jones.com , book.)

Is the universe a computer?

"What is computation?" was the subject of a 2010 ACM symposium and "what is the nature of reality?", the universe, is an open question in philosophy. In my blog of 28 Aug. 2012 I suggested that a computer might best be described as a reconfigurable causal network. In the classical physics clockwork view of reality the universe certainly is a causal network.  The problem is its not all that reconfigurable.  The big bang did the configuring and that's hard to change. Perhaps equate a multiverse with a computer then. Provided there are a variety of initial conditions.
 If you mean by "the universe" only that portion of reality which is external to yourself then your actions might do a tiny bit of reconfiguring, but not much. You plus the universe would be a coupled pair of computers/Turing machines. I run some Asa H simulations in this way. (See my blog of 10 March 2016.)
One can generalize these notions to a quantum computer and a quantum mechanical reality/universe.

Reconceptualizing reality

As evolution changes the uses of things, both physical things and mental things, it also necessarily changes their meanings.  So ontologies must change over time.  Usually these are small and gradual changes.  Occasionally they are the more profound changes I have been thinking about like the transition from classical physics to quantum mechanics or relativity theory.

It From Bit?

If I am serious about the possibility of reconceptualizing reality I should be willing to at least consider John Wheeler's It from Bit postulate. To do this I am reading Aguirre and Foster's It From Bit or Bit from It, Springer, 2015.

Actually, the history of a computational reality notion dates back to Konrad Zuse. Zuse called it the computing universe. See, The Computer - My Life, Springer-Verlag, 1993, pg 175.

If all we really experience is Hume's bundle of perceptions or Lewis' observational terms it might make sense to ground our ontology on information. One could stop there and call it idealism?

Thalos has attacked the idea that we should be trying to describe everything on some one single lowest level of reality (See, Without Hierarchy, M. Thalos, Oxford Univ. Press, 2013).  Others suggest there is not one single "correct" or "best" ontology. (See, The Logic of Reliable Inquiry, K. Kelly, Oxford Univ. Press, 1996 and Constructing the World, D. Chalmers, Oxford Univ. Press, 2012) This is in line with scientific pluralism.

Asa H's ontology seems to evolve over time and depend upon the order in which it receives its experiences.

What is the "best" ontology will depend on what I'm using it for.  Quantum fields might be the best description of ultimate reality but they would not be the best vocabulary with which to talk with friends in daily life.  They might not be the best basis for a private language or language of thought either. This was a reason for adopting scientific pluralism. Even in a single branch of science one might find use for several different "languages." The Heisenberg picture of quantum mechanics versus the Schrodinger picture for example. Are wave functions a function of time?  Are operators a function of time? 

It would seem that we need several different (but possibly overlapping) ontologies. And I expect these to change over time.

Actions

Is all that we can "directly know" just our sense impressions (input signals)? That is, Lewis' "observational terms", O-terms. Hume said "...only the successive perceptions constitute the mind." (A Treatise on Human Nature, 1739) Asa H does learn just such sequences of sensations.  But for Asa H as well as for humans there are also actions (output signals).  And these are important for optimization/intelligence. Simple observation versus full blown experiment.  In Asa H extrapolations, deductions, etc. serve as hypotheses for future testing and correcting.

Complexity and diversity

I am reading the book Complexity and the Arrow of Time by Lineweaver, et al. (Cambridge Univ. Press, 2013)  Perhaps the problem is that complexity and diversity are again simply the names of vector quantities, quantities having components like: genome length, the number of cell types, the number of niches, the number of species, the specialization of body parts, the number of component functions, etc. (If these things can be measured.)

Asa H and associationism

At a conference a few weeks ago a colleague suggested to me that Asa was a software implementation of Hume's associationism.  While I agreed in part I did point out that a lot more was also going on in Asa H besides association.  See, for example, my website www.robert-w-jones.com, cognitive scientist, theory of thought and mind. I referred him instead to John H. Andreae's work.

 But the Asa H experiments ARE relevant to philosophy (as described in many of my blogs).  According to David Hume the "self" is "...a bundle or collection of different perceptions which succeed each other..."  This is exactly what Asa's concept of its self is as I've described while it's been evolving.

Asa H discovers mind-body dualism

My artificial intelligence Asa H has formed a concept composed of the actions it is able to perform. This concept is, in turn, composed out of what might be termed a concept of mental action and a separate concept of physical action.  The concept physical actions is composed of things like moving, turning, grasping, lifting, etc., all actions that require substantial current draws on the NXT/EV3 batteries.  The concept mental actions is composed of things like extrapolation from the case base, searching through the case base, loading a data (case) file, sorting a file, etc., all actions that involve no large added current drains. (To increase its utility measures Asa prefers to take actions that do not require a large current drain.) Asa H makes this distinction between the mental and the physical.

Consciousness in Asa H

Typical Asa H light software (see my blogs of 10 Feb. 2011 and 14 May 2012) allows for simple adjustments to learning by setting parameters like L and skip. More complex software packages allow Asa to observe the amount of time it spends taking input, giving output, searching the case base, performing feature extraction, adding to memory, sorting memory, comparing, extrapolating, doing deduction, doing simulation, case updating, etc. and then correlate these efforts with the utility (rewards) observed/received over time (see chapter 0ne of my book Twelve Papers, the section titled self monitoring  www.robert-w-jones.com). Parameters like L and skip are, themselves, made inputs to the hierarchical memory and Asa learns a vector/concept like:

 thought = (search, deduction, simulation, sorting, extrapolating, comparing, remembering, etc.).

 Asa can be allowed to adjust the learning itself by making the parameters outputs of the memory hierarchy. Thinking can come to constitute a part of Asa's concept of its self:

 self = (sense, act, health, thought).

This is a further evolution of Asa's self concept. Asa can observe some of its own thought processes.

 In interaction with the world I have tried to give Asa the same sort of sensations and behaviors that a human might experience.  If Wittgenstein is right this might be necessary if humans and AIs are to understand each other. But what Asa sees of its own thought processes is quite different from what humans know of their own inner thoughts.  Will this prove to be a problem? Might the same thing be true if we met space aliens?

One, two, three and four dimensional memories for Asa H

In my blog of 7 Jan. 2015 I describe how to give my artificial intelligence Asa H a one dimensional memory for things like recorded speech and a two dimensional memory for things like images.  With Asa H now controlling a distributed set of Lego NXT and Ev3 robots it is also possible to establish a three dimensional memory with the agents distributed about in 3 space. Since this is recorded as a function of time it is a four dimensional pattern in memory.

This hardware and software configuration quickly forms the "action-at-a-distance" concept as it learns.

There is not one single fundamental level of reality

Some empiricists might tend to believe that there is a single most fundamental level of reality be it strings, or quantum fields, or what have you.  Schaffer has argued against this view (Nous, 37:3, pg 498, 2003). Concepts like Chalmers' indexical I (Constructing the World, Oxford U. Press, 2012, pg 390) or Wierzbicka's substantive I would correspond to the self concept that my AI Asa H is learning (see blogs of 5 Dec. 2015 and 4 March 2015). This concept resides on a fairly high level in the Asa H case memory hierarchy. Other concepts like Chalmers' quality color or Wierzbicka's touch or hear reside at much lower levels in Asa's hierarchical memory. Furthermore, I do not think we need to believe equally strongly in all of our concepts, even our most fundamental ones.  (Scientific pluralism again.)

Seaking ultimate O-terms

In his Aufbau project Carnap argued that all concepts could be constructed from a similarity relation and a few logical concepts (like AND, OR, NOT). These would be, in effect, Lewis' ultimate O-terms.  Asa H has (one or more) similarity measures and NOT built in (innate) and can learn sequences that implement AND or OR.

In addition to Anna Wierzbicka's 63 semantic primes Locke offered 8 ultimate conceptual primitives: extension, solidity, mobility,perception, motivity, existence, duration, and number (Essay Concerning Human Understanding, book2, chapter 21, 1690). But some of these can be decomposed into other primitive concepts (more primitive ones?). Solidity can be learned as a sequence of actions involving touch, force or pressure application, and observed degree of deformation/deflection.  Mobility can be learned as a sequence of detecting, touching, grasping, lifting, and carrying. (Note that outputs, actions, as well as inputs, sensations, are involved.)

In this way we can try to find a more primitive (most primitive?) set of O-terms.  Or, by recombining the basic conceptual elements (subelements?) could we hope to operationalize our project to reconceptualize reality? Physicists, for example, might wish to combine Locke's perception and existence into one single concept.  The idea being that what exists is whatever can be detected/measured by sensors/instruments.

Empiricists like Lawrence Barsalou and Jesse Prinz believe that all the most primitive concepts are acquired by direct perception alone. (Furnishing the Mind, Bradford Book, 2004) In that case intelligences with the same senses might expect to share the same (or at least similar) fundamental concepts. We might expect Asa H to reconceptualize reality due to its ability to perform radio ESP, echolocation and ranging and to directly sense GPS location, electric fields, magnetic fields, atmospheric pressure,  and nuclear radiation.

Impact factors and citation analysis

I was talking with a college administrator at a conference this weekend and the discussion turned to measuring research quality. While I don't think they are totally useless I do believe that, like student evaluations of instruction, citation analysis of scientific research is largely a popularity contest.

I want the truth of a scientific idea to be measured by its agreement with observation and experiment. I don't believe science is democratic.  Human opinion can not be the deciding factor.

What about the most difficult of fields where only a handful of people are even capable of understanding the concepts involved? (Feynman said that nobody understands quantum mechanics.) There won't be many papers to cite or readers to have read them. Popular is not the same as good.