Thursday, July 26, 2018

Matches and lighters

Maybe we shouldn’t compare reusable rockets with airplanes. Reusable rockets need parachutes, wings, extra fuel, etc. and added design and maintenance costs. Their economics depends upon the number of launchs that can be expected over time. Just like we still manufacture and use both lighters and matches there may be room for both reusable and non reusable rockets in the world’s fleet. Certainly military rockets like ICBMs will be single use. And when they are decommissioned why not use them as space launchers as we have in the past?

Tuesday, July 24, 2018

Embodiment

Two different A.s.a. agents will have learned somewhat different concept webs. These webs will differ even more if the agents are specialists of different types. This makes it harder for one agent to tell another agent what it knows or simply what it sees. On the other hand, I can understand and make use of physics theory and data obtained from experiments I have never performed myself.* I only need to have some somewhat similar experience behind me. Must every AI be embodied at least to some degree? If so, to what extent? Or, can enough be simply “disc copied” or hand coded into any unembodied agents? With A.s.a. I have done this successfully, but with small scale “toy” examples only.
* And some of us are blind. And some of us are deaf. And some of us are paralyzed. .......

Incorrigible ontological relations

Julian Galvez argues that the human mind comes to model/understand the world by application of the primitives: difference, similarity, property, and causality. (Our Incorrigible Ontological Relations and Categories of Being: Causal and Limiting Factors of Objective Knowledge, 2016) A.s.a. H. uses the vector dot product and/or other similarity measure to obtain similarity and difference assessments. A.s.a.'s construction of a concept hierarchy defines properties. A.s.a.'s sequence learning covers causality. A.s.a. also does things like averaging over multiple observations, however.

Saturday, July 21, 2018

Syllabi for AI training again

For many years I have stressed the importance of the syllabi I need in order to be able to successfully train an AI. (It was included in my 9 Sept. 2010 blog.) Alexander Wissner-Gross attributes progress in AI to the use/availability of high quality data sets. (See edge.org/response-detail/26587)

Wednesday, July 18, 2018

Will any sufficiently intelligent system exhibit consciousness?

I have argued that A.s.a. H. exhibits machine consciousness. (See, for example, my blogs of 21 July and 19 October 2016.) The state of anything like a Moore or Mealy machine will develop sensitivity to unique temporal sequences of inputs and David Hume's view of consciousness was as "...a bundle or collection of different perceptions which succeed each other..." (A Treatise on Human Nature). For more demanding definitions of consciousness things are not so clear. Many embodied AIs might sense damage ("pain") and the need to recharge batteries ("hunger") and so exhibit Ned Block's "P-conscious" states.  Block's "A-conscious" states, things like "grass" having the feature "green" might depend upon how the AI organizes its knowledge base. Metacognition and Hobson's functional components might also be at issue. (Things like emotion.)

Monday, July 16, 2018

Degrees of realness

Luciano Floridi has argued that our "reality is the totality of" our "information." (The Philosophy of Information, Oxford University Press, 2011, page xiii)  If we were to employ that definition of realness then not all things need be equally real. How real a concept is would depend upon how many models/theories it appears in, how important a role it plays,  and how strongly and how frequently that particular concept is activated during cognition. The quantum wave function, for example, is quite real in David Albert’s version of Bohmian quantum mechanics and not at all real in Bradley Monton's interpretation of quantum mechanics, when he says: "The wave function, according to Bell, is an inessential mathematical device...". (See The Wave Function, Oxford University Press, 2013, pages 108 and 162) Different definitions of what realness is would also have an impact.

Tuesday, July 10, 2018

Science versus capitalism

The current issue of Physics Today has an article on how and why business is working to keep their research ideas and results secret. (Douglas O'Reagan, “Who Owns a Scientist’s Mind?” Physics Today, July 2018, pg 43) Science, on the other hand, operates best when we all share our results and methods openly.

Sunday, July 8, 2018

Natural intelligent system

A.s.a. H. Is a project to engineer general intelligence. It was not biologically inspired. But it is possible to employ Grossberg’s A.R.T. networks as the clustering modules in A.s.a. H., and A.R.T., adaptive resonance theory, is biologically plausible giving us a biologically plausible version of A.s.a. H.

Friday, July 6, 2018

Credit propagation

Once A.s.a. H. Has created a substantial hierarchical model of itself and its environment it is possible to perform sensitivity analysis starting at the top of the hierarchy* and working down. Whether  scalar or vector utilities are employed the question becomes how to weight the credit(s) computed at each successively lower level in the hierarchy.

* Starting with the utility(s) measured for the complete agent.

Monday, July 2, 2018

The curse of dimensionality

The natural world is a very high-dimensional state and action space. Some of the ways I have tried to deal with this complexity are (in no special order):

Hierarchical decomposition/learning (A.s.a. H.)
Clustering
Approximation
Forgetting
Ordered, organized training syllabi
Hand coding and human supplied problem solutions
Attention mechanisms
Multi-agent/specialist AIs
Parallel processing
Pre and post processors

in the future we may have available fast quantum computers which would also help