Growing Asa H's concept of its self

In my blog of  4 March 2015 I identified a fragment of Asa's hierarchical case vector memory which constitutes an initial concept of self. I am watching this concept grow as Asa continues to interact with its environment.   Asa has added (or modified) the vectors (concepts):

push=(move to, touch, feel contact force)
kick=(ball near, push, ball far)
self=(health, grasp, kick)

What concepts are learned clearly depends upon the bot's detailed anatomy, sensors, and actuators. In a different world quite different concepts evolve.

Asa H's language of thought

On each level in Asa H's hierarchical memory Asa defines and evolves symbols/words/concepts/categories.  Patterns are developed between levels which link these concepts into a semantic network.  This system is Asa's language of thought.  It differs from typical human languages in the degree to which it is hierarchically structured.

Constructed memory

Like humans, Asa H has a constructed memory. Typically, there is filtering, weak inputs are not retained at all.  When retained, an averaging is usually performed (with previous very similar memories).  Only the more strongly activated cases pass on activation to the next higher level in the Asa hierarchy.  Forgetting may be used in order to clear/maintain space in (limited) memory.

A forgetting heuristic

Retain longer those memories/cases with the highest and lowest utilities.

Work on automatic programming

At a conference a few weeks ago I was asked if I had done any work on automatic programming.  That set me thinking.  I have done a little with genetic programming, but really very little.  I suppose some of my neural network work (and, for that matter, Asa H work) could be viewed as automatic program generation from data/examples.  As I think about it, however, the most practical work I've done along these lines is probably the assembly and subsequent use of my code library.  Most of this can be viewed as a component library.

Asa H as an informal system

Formal systems "have to fix the language and the rules of operating on symbols beforehand....to definitely exclude subjectivity"  (see Problem solving with neural networks, Wolfram Menzel, Institut fur Logik, Univ. Karlsruhe, Germany)  Asa H, then, is an automated INformal system, a system that defines and then refines its symbols and language on each of the levels of its hierarchy.  The rules of operation that apply between levels of the hierarchy are also variable over time.

Useful science, useless science

I have commented before on the analysis that suggests that the majority of scientific publications are, in fact, wrong. (see New Scientist, 30 Aug. 2005)  (I'm sure that the exact proportion varies some from one scientific field to another.)

I have observed that there are also a lot of papers that may not be wrong as such but which are just not very useful.  I won't name names but there is, for instance, a lot of work in computer science that involves the same old methods and algorithms but rewritten in whatever programming language happens to be popular at that time.

The reality of the wave function or quantum fields

I have argued before that not all ontological entities are equally real.  If what is real in the world is what has explanatory usefulness then not everything is equally real.  Not all concepts/entities have equal usefulness.  Deutsch and then Wallace (The Emergent Multiverse, Oxford Univ. Press, 2012, page 389) argue that there are not enough atoms in the universe to account for how Shor's quantum algorithm can factorize a number. The required machinery must be seen to be in the form of quantum fields, not matter. This argues strongly for the reality of the high dimensionality quantum realm. (But I would not necessarily say this has to be in the form of a set of emergent, non interacting,  nearly classical, Everettian worlds.)

Hierarchy of laws

In cognitive science operations in the cognitive or "knowledge level" are performed by lower level components of the program level.  Operations in the program level are, in turn, performed by components of the register level.  This decomposition continues from the register level down through the logic level, circuit level, and device level.  Each level has its own laws of operation. (Unified Theories of Cognition, Allen Newell, Harvard Univ. Press, 1990)  The program level, for instance, is typically composed of sequences, decisions, and loops.  The circuit level, on the other hand, is governed by Ohm's and Kirchhoff's laws.  In human beings the circuit level would be replaced by a network of neurons and the device level would describe those neurons using laws of electrochemistry.

Laws which are valid when applied to one level in this hierarchy may be invalid if applied to another level.  Boolean or propositional logic is valid in a computer at the logic level.  But in Asa H, a nonstandard logic, or fuzzy logic program different laws of logic are valid at the program level. A PC running a simulation of a quantum computer might be another good example.  The simulation is following quantum laws while the PC is following classical laws.  Philosophy of mind has sometimes erred by trying to apply the wrong laws to the wrong level.

Asa H (and any other intelligence) in turn builds its own cognitive levels on top of these.  As concepts like hunger/need, obstacle, damage, health, danger, and self evolve so too do notions of agency, good and bad, etc. Laws/rules that apply to these cognitive levels may not apply to other levels in the hierarchy.  (Things like social norms, morality, and the like.) In general, regularities/patterns (i.e. "laws") that are exhibited at one level of detail/abstraction, and with one set of concepts, may not be found on other levels.

If there are "laws of thought" then these would be valid in one or more of these cognitive levels.  Might something like "free will" or "moral responsibility" be a reasonable description in one of these  cognitive levels but not elsewhere? (see my blog of 21 Jan. 2015)  Pluralist science again.

What is "real" in the world is what has explanatory usefulness.  It may be useful to attribute "free will" to a person.  Rather than acting in the way s/he has previously acted s/he might do something different and we might want to be prepared for that today.  If the person has acted "morally" in the past s/he may likely act "morally" today also.  At some other level of description "free will" or "morality" may be useless  (invalid) concepts.  Newton's laws are valid on a level describing the macroscopic world.  They are invalid when describing the microscopic.

AI, specialization, and reductionism

Much of science has been built following reductionism and specialization.  This has been criticised in artificial intelligence (see, for example Artificial General Intelligence, Goertzel and Pennachin, eds., Springer 2007),"narrow AI". But in creating my Asa H AI I have found nearly all of the AI subfields helpful.  (for a list see my blog of  9 Sept. 2010)  Asa H has made use of algorithms, concepts, and methods from most all of these specialties. Specialization ("narrow AI") seems to have proven quite fruitful.

AI data warehousing

I have an extensive code library (see blog of 20 Feb. 2014) that contains algorithms from all of the major AI subfields (see blog of 9 Sept. 2010). What I need now is a data warehouse covering the AI curriculum I've discussed in blogs of 18 July 2014 and 23 March 2015.  The machine learning community has, for example, the 131 data sets in the UCI repository, DMOZ, etc.  Organized in order of complexity I currently have data sets for logic functions, numerals, letters, words, phrases, pictures, and movies.

Vector values again

In discussing the judging of student papers at the Kansas academy of science conference it was again obvious to me that quality can not be reduced down to a single scalar value.  Scientific papers must be judged in a 3 or 4 dimensional space.  One dimension involves the quality of the data taken or the experiment performed.  A second scale involves the quality of the analysis or theoretical work done.
A third dimension would measure how creative or original the work is.  A possible fourth scale might measure the amount or volume of work done.  If a given paper is best in all of these measures it can be ranked as best overall.  But otherwise there is no "best."

Administrators

Years ago I taught a summer course to 2 students.  It was considered important.
Another year my physics course didn't run but they found me an algebra course to teach.
These days they have no work for me in summer.  Another question of values.

Values, learning, and science

I have argued previously that science, like other forms of cognitive processing, can not be valuefree.

A reinforcement learner accepts among its inputs a stream of rewards.  These constitute some of what it values.

A learning system like a backpropagation neural network may have no reward stream but if it learns from sets of inputs and outputs one of the things it will slowly learn is preferences (values) inherent in the training set supplied to it.

A learning system trained on input (observations) alone (no output actions) will  value things like amount learned, speed of learning, precision of recall, etc.  These will be built into the learner in the form of various thresholds, learning rate parameters, vigilance parameters, similarity measures, etc.

Science will not be value free any more than any other cognitive process going on in such a machine/human.