Sunday, October 26, 2014

Parallel computing

Parallel computing is especially useful for neural network training.  I have trained some neural nets for as long as a week.  I have trained 3 or 4 of these at one time on 3 or 4 conventional computers.  I have enough computers that I could train at least 10 neural networks at a time, giving a factor of 10 speedup. This is useful for training preprocessor networks, expert neural networks, voting neural network classifiers, modular neural networks, etc.

spatial and temporal patterns

My Asa H 2.0 artificial intelligence receives a stream of inputs and generates two output streams; one composed of physical actions taken in the world and the other a set of models that describe the world Asa finds itself in.  Of all the patterns Asa comes to recognize in its input stream the majority are spatial patterns.  Only a minority of learned input patterns are temporal patterns.  Of all the physical outputs Asa learns the majority are temporal patterns.  Only a minority of the physical outputs learned by Asa are spatial patterns. See also my blog of 22 Sept. 2014.

Friday, October 24, 2014

A danger in religion

If you believe in souls and life after death you may allow yourself to do things that are extremely dangerous for you and for society.

Wednesday, October 22, 2014

Counting

I have taught Asa H 2.0 to count (small numbers).

Vector values

In Asa H 2.0 light I typically use pattern length and frequency of pattern occurrence as components of a  case's vector value/utility. (see my blog of 19 Feb. 2011) A pattern's complexity can also be used in measuring its importance.  It is not clear which measure of complexity to use, however, permutation entropy? (Brandt and Pompe, Phys. Rev. Lett., 11 April 2002) Ke and Tong's measure? (Phys. Rev. E, 2008) or what?

Diversity in a society of Asa agents

The various agents can be trained in a wide variety of specialties.  The agents may have different values; one agent may value lifespan more than offspring.  Another agent may value offspring more than lifespan.  One knowledgebase may contain cases that value case length or complexity more.  Another knowledgebase may contain cases that value frequency of pattern recurrence more.  Such diversity will help the society deal with complex time varying environments.  The society of agents will be more capable than a single agent.

Sunday, October 19, 2014

Asa's fuzzy protologic

One sort of proto-logic works by verifying if a subset of symbols is present in a certain set. (Principles of Quantum Artificial Intelligence, Andreas Wichert, World Scientific, 2014, pg 31 ).  The set is represented by a vector which is divided into sub-vectors.   Asa searches for sub-vectors in this way but it is satisfied with an approximate rather than an exact match.

Saturday, October 18, 2014

Collective leadership

Having multiple models of the world is better than having just one (see my blogs of 17 Aug. 2012 and 13 Aug. 2012).  As a consequence diverse groups make better decisions than individuals do (and I dislike traditional  managers, governors, presidents, etc.).  A society of Asa agents may outperform an individual one.

Friday, October 17, 2014

Naming robot body parts

The concepts, "touch left", "touch right", "touch front", and "touch back" have been given to a Lego NXT robot by touching sensors at those locations in association with language input of the respective terms. The more general concept of touch is then learned at the next higher level in the hierarchy.  One can similarly locate and name any number of body parts that have input sensors.  Internal organs like the battery and recharging circuit can be identified with the sensation of "charging", "high charge", and "low charge." Categories like "right" can also be learned at a higher level in the hierarchy by association of experiences with "touch right", "turn right", etc.

Wednesday, October 15, 2014

Possible conscious states in Asa H

There are many different theories of consciousness and I have discussed these before in connection with Asa H and AI in general.  I still believe that there is something to a number of these different theories. I just want to add an additional observation from my work with Asa H.

As Asa receives a stream of input it may identify one or more cases (categories or patterns or sequences) in its case memories that closely match this stream (so far at least).  Any actions that will be triggered are predictions taken from these remembered cases.  In order to reduce search (when one is considering a large, complex case base memory) I sometimes stick with the same active case(s) so long as the degree of match with new input does not degrade too much.  (i.e., I don't do a search through all of the case memory every time a small new amount of input appears.)  The cases which are present in the memory but which are not active and not being searched through at that moment might be considered to be in Asa's "unconscious."  The case(s) that are active, that Asa is following at this moment, might be considered to be what Asa is "conscious" of.  Typically the amount of memory that Asa is conscious of is much smaller than what Asa is unconscious of.  Consciousness here results from the need to reduce search/computational complexity. Any actions that are triggered come from the conscious case(s).

We might choose to keep a maximum of 7-9 cases active.  (But at how many levels in the hierarchy?) The unconscious cases are in a long term memory.  The conscious cases are held active short term. Substantial thought (the Asa extrapolation routine for example) is running subconsciously (and is creative).

Tabula rasa learning again

I'm going to try the following; watch as a layer (in Asa H) learns cases/categories.  If the total number of categories being learned levels off, at that point only, begin to train the next layer up. Repeat for each layer on up.  I am assuming all of the input has been structured, presenting simplest patterns/categories first.  That may be a lot of work.

Tuesday, October 14, 2014

Boxology

Sometimes we know how to implement the boxes we draw:



But all too often in artificial intelligence/cognitive architecture work the boxes are opaque.  We have little or no idea what to put in them:



With Asa H I hope I have made clear (at least some ways in which) you can implement the boxes. (of my diagram at www.robert-w-jones.com , cognitive scientist, theory of thought and mind)

 I went on and  offered some actual code in my blogs of 10 Feb. 2011 and 14 May 2012.

Tabula rasa Asa H

When training Asa H starting from an empty case-base I typically present the simplest/smallest things (categories) first, training Asa much like we might teach a human infant.  Perhaps we should also turn off/postpone learning in the higher layers of the Asa hierarchy until the lowest layer(s) have learned the simplest/smaller categories. How long should this delay be for each layer?

Monday, October 6, 2014

Values

The objective of life is to survive, expand, and fill as much of space and time as possible.  Intelligences inherit this goal and values. A vector value would then include the agent's lifespan and the sum of the lifespans of all of the agent's offspring.  A measure of the spread of the offspring over space might also be included in this. During the agent's life it will only be able to estimate its ultimate lifespan.  Even if offspring only have one parent they will learn and evolve during their lifetimes and  only partially reflect the value/utility of their parent. Offspring lifespans will also vary.  The simple scalar utility, U=L(1+N), where L is the agent's lifespan and N is the number of offspring, is then an even cruder estimation.  When cooperative agent societies are considered, diversity, in the form of a wide range of agent specialists, should also be valued.

Thursday, October 2, 2014

Hard-wiring Asa H

It may be possible to speed up Asa H by hard-wiring at least some of its functionality.  As an initial step in this direction I am loading a portion of Asa's case memory onto FPGAs.

Lone wolf

I have tended to work alone.  This is partly  my personality.  I have also been willing to change directions "on a dime."  You can't do that if you work in a group. Neither have I wanted to be confined to working on purely "mainstream" topics.  I've wanted greater freedom than that allows.

Wednesday, October 1, 2014

Love

Some religious thinkers and philosophers take love to be a foundational quantity.  Rather, a biological theory of love would take it to be an emergent quantity, evolved to aid in survival of offspring, family bonding, or societal bonding and cooperation.  It might be one of the primitive drives which form a part of the (imperfect) human value system.