Saturday, October 29, 2016

Dynamic adaptive case libraries

It would certainly be fair to call Asa H's hierarchical memory a dynamic adaptive case library (DACL). In his new book A Dynamic Adaptive Framework for Case-Based Reasoning (LAP Lambert, 2016) Orduna-Cabrera outlines some original ideas on DACLs including indexing structures and methods. I may try to adapt some of these for use in Asa.

Thursday, October 27, 2016

Another argument for alternate realities

In his book The Philosophy of Information (OUP, 2011) Luciano Floridi argues that "...reality is the totality of information..." possessed by a "semantic engine", i.e. a human or an AI. (page xiii) Since different people and different AIs will possess somewhat different information/knowledge bases they will experience somewhat different realities.

A.s.a. H. has learned to lie

My artificial intelligence Asa H has learned how to lie. Some of the first words I taught Asa were things like animal warning calls, calls that cause other agents to flee and/or hide or calls for other agents to come and help. (Help push an obstacle out of the way for example.) Asa learns that these calls (calls to other Asa agents in a society of agents or calls to a human nearby) result in the other agent(s) fleeing or approaching. Asa comes to learn that if it wants to "hog" a "food" source (battery recharging station) or other limited resource it can issue a warning call, a warning of danger that doesn't actually exist.

Wednesday, October 26, 2016

A moral machine

A.s.a. learns to predict and optimize/improve lifespan and diskcopy/reproduction at the top of the case memory hierarchy, Pains (or NOT(pain)s), battery charge, and component malfunctionings (things like motor stalls and subsystem failures) are seen/measured as inputs to the bottom of the memory hierarchy. Pattern size/length and frequency of occurrence are seen as values on any and all levels of the memory hierarchy. For A.s.a., improvement, as measured by increases in these values, constitutes its moral calculus. We are not intending A.s.a. to be a servant.

Guns on campus

On 18 October 2016 the Emporia State University faculty senate passed my resolution FSR 16003:

Whereas, firearms make it easier to kill people, and
Whereas, we wish to make it harder to kill people, and
Whereas, there has been too much gun violence in our society already
Therefore, be it resolved that all firearms should be prohibited on the campus of Emporia State University.

Monday, October 24, 2016

Rigged elections

The GOP should know all about rigging elections. They stopped the vote counting in Florida in 2000, the GOP controlled supreme court picked the president in 2000 (though Gore won the popular vote), they have gerrymandered control of the house, and they suppress voting where ever and when ever they can.

Wednesday, October 19, 2016

Machine consciousness

This is a draft of an abstract for a presentation on machine consciousness that I am working on:

Isn't Spacex trying to do too many things at once?

ISS resupply, F-9 reuse, Falcon heavy, manned Dragon, ITS.
It seems to me that if you try to do too many things at once either they all come in late or some of them fail or both.

This is an update of my 30 September 2011 blog.

Friday, October 14, 2016

Attention as response

One approach to handling/modeling attention, or at least one kind of attention, is to treat it as a response, either an innate response or a learned one. Things like turning toward a stimulus, increasing  the gain on a microphone, adjusting vision magnification, turning on and bringing to bear additional sensors, etc.

Thursday, October 13, 2016

How human-like should a robot be?

There are those who argue that if a robot is made as human-like as possible this will help the two relate to each other and understand each other. See, How to Build an Android by David Dufty, 2012 and Virtually Human by Martine Rothblatt, 2014. I tend to disagree. I think this just makes the robots creepy and harder to relate to. But it is true that body form and function influences the concepts the robot will develop and use. That will aid in robot-human understanding. I think something along the lines of  Softbank robotics' Pepper is a reasonable compromise.

Wednesday, October 12, 2016

immortality

I have argued that immortality is impossible. (see my blog of 15 October 2010) I had expected, however, that there was room for a considerable increase in human lifespan.  But Michael Ramscar of the University of Tubingen says that even at our current ages "Some things that might seem frustrating as we grow older are a function of the amount of stuff we have to sift through...and are not necessarily a sign of a failing mind. A lot of what is currently called decline is simply learning." (see The Myth of Cognitive Decline in Topics in Cognitive Science, 6, 2014, 5-42) Or, as Christian and Griffiths put it "what we call cognitive decline may not be about the search process slowing or deteriorating but at least partly an unavoidable consequence of the amount of information we have to navigate getting bigger and bigger." (Algorithms to Live By, Holt and Co., 2016, page103) I am not arguing that there are not things like alzheimers (my mother died with it).  What I am arguing is that it may not be possible to have the kind of immortality some people hope for.

Monday, October 10, 2016

AAPT conference

At the American association of physics teachers conference this past weekend James Laverty of Kansas State University presented the 3D-LAP scheme (3 dimensional learning assessment protocol) for assessing the value/importance of various physics test questions.  I noted favorably that the method employs a 3 dimension vector value:
1. Scientific and engineering practice
2. Cross cutting concepts
3. Disciplinary core ideas
I am not sure these 3 are exactly what I would have come up with but I am obviously in favor of vector value systems in general.
I also was interested in the scientific and engineering practices they identify (from A Framework for K-12 Science Education):
1. Asking questions and defining problems
2. Developing and using models
3. Planning and carrying out investigations
4. Analyzing and interpreting data
5. Using mathematical and computational thinking
6. Constructing explanations and designing solutions
7. Engaging in argument from evidence
8. Obtaining, evaluating, and communicating information
Since I believe that the process of science is simply the process of intelligent thought (perhaps refined and augmented in various ways) these are then all things that my artificial intelligence A.s.a. H. should be doing too. Said another way, Asa should be able to do science.
1. Asa defines and identifies cases that lead to low utility, i.e., problems.
2. Asa's hierarchical memory creates, stores, and uses spatiotemporal patterns, models of reality.
3. Asa examines the accuracy of its extrapolations experimentally and plans future behaviors.
4. Asa examines its case memory using interpolation, extrapolation, value assessment, etc.
5. Asa is computational and uses mathematical as well as logical reasoning methods.
6. Asa designs improved behaviors to cope with problems, i.e., low utility situations.
7. Asa reasons ("argues") from evidence.
8. Asa can communicate and output its case memory.
I would like to improve upon Asa's present ability to ask and answer questions.

Friday, October 7, 2016

Attention in AI

I remain dissatisfied with our ability to focus attention.  I believe that this will become more and more of an issue as we scale up applications of AI.  One idea that might be useful is the concept of thinking about something in the right way.  Perhaps specialist AIs can be built around groups/clusters of specialized concepts, knowledge, and algorithms. What would the right clusters be?  How would we modify/learn them over time, perhaps dependent upon environment?

Thursday, October 6, 2016

The origin of consciousness

Although A.s.a. H. was not biologically inspired I do believe that consciousness in animals may have developed in the same sort of way that a sense of self develops in A.s.a. (As described in my recent blog posts, especially 1 October and 5 November 2015 and 21 July 2016.)

Some people have questioned why nature would evolve consciousness.  In Asa H the usefulness and survival value of consciousness is directly observable.


Tuesday, October 4, 2016

Complex levels of reality

In my blogs of  13 April 2015 and 12 April and 18 June 2016 I have argued that there is not one single fundamental level of reality.  At least not as described in our current best models. Rather, there are multiple levels. Things that are true at one level of description may not be true at another.  In the macroscopic world things may be wet or dry, I can measure this property.  Asa H measures the humidity of the surrounding air for example.  But things in the microworld are not wet or dry. It makes no sense to try to measure if an electron is wet or dry for example. That property doesn't exist at that scale. It's not a relevant concept there. Spatiotemporal patterns that are found on one level of description may, or may not, be found on other levels. Reality is then described by the concepts, patterns, and laws taken collectively from all levels. (See, also, The Philosophy of Niels Bohr, H. J. Folse, North Holland, 1985. On page 166, for example, Folse describes how the concept of temperature would have been valid all the way down if classical laws had remained true at all scale sizes.)

Saturday, October 1, 2016

GPU computing

The Asa H project has always involved a certain amount of work on parallel processing. Each of the various levels in the case memory hierarchy could be running on a different computer, for example. (e.g., my blog of 14 December 2015) I am now looking at CUDA C/C++ to see if I can incorporate parallel processing with GPUs. (using a NVIDIA GeForce card) GPU computing has been used for machine learning, pattern matching, feature detection, and speech recognition for example.