Monday, September 29, 2014

A convolutional Asa H network

By using multiple copies of the Asa H 2.0 code in each layer of the hierarchy and by transferring copies of learned cases between these (code copies) it is then possible to construct a convolutional network out of Asa H.  Convolutional networks have proven to be useful in object recognition, for instance. (see the work of Y. LeCun)

Saturday, September 27, 2014

Thursday, September 25, 2014

Asa H action selection

At any given time various layers of Asa H may predict 1 or more cases that will be active next.  Higher layers make their predictions based upon a longer span of inputs.  (And make a larger number of predictions that reach further into the future.) Some of these predicted cases can involve predicted actions Asa could perform. During the course of experimenting with several hundred Asa H programs I have tried out a number of different action selection algorithms.  With the entire Asa network acting as an evaluation function each of the possible actions can be "tried" (simulated) to see which gives the highest utility. No output action is actually taken during this simulation/evaluation stage.  The simulated action, actions, or non action which gives the highest utility (measured at the top of the network hierarchy) is then selected and is scheduled to be taken 1 or more time steps into the future. (Actions scheduled far enough into the future could subsequently be preempted by further future predictions and evaluations.) To reduce search/complexity one can put a time limit on how far into the future one looks and tries to predict (establish an "horizon").

Wednesday, September 24, 2014

Asa H vs other deep learners

Other deep learners typically have a fixed number of layers, a fixed number of nodes, a fixed number of nodes per given layer, etc.  Geoffrey Hinton's ImageNet, for instance, had 7 layers, 650,000 nodes, total, and a fixed number of nodes in each given layer.  Asa H, on the other hand, adds layers and cases/concepts as it learns, and the number of cases per layer varies with time as Asa learns.

It is an advantage of Asa H over humans that it can add memory and processors as and when it needs them.

Tuesday, September 23, 2014

Immorality

Immorality appears to be a vector having components like:
1. harmfulness
2. dishonesty
3. disgust

Monday, September 22, 2014

Asa H output

Our experiments with Asa H have involved much more input than output.  Humans have a couple hundred million rods in their eyes and only a few hundred skeletal muscles.  A scientist might be able to read perhaps 100 papers for every one he writes himself.  An intelligence in the world simply gets a lot more input than the output it generates. In the lowest layers of the Asa hierarchy overt action is rare.  I typically mark those few cases that involve action in order to find them easily.
I am also especially interested in concept formation by AIs (how conceptual knowledge is formed from perceptual inputs) and have spent some time studying that. Such concepts are a kind of "output" but most of my robots can only do what Lego NXT servomotors can do. I would like to upgrade this as well as my input sensors.

Thursday, September 18, 2014

Magical thinking

Is the relationship between math and physics descriptive or prescriptive? I view math as a language used to describe what I see and as a theory (or a set of theories) of patterns. At one time people thought that gods had secret names.  If you knew the name you could call up a god to do your bidding. Abracadabra. Open sesame.  Today some people think mathematical "rules" force nature to be some way, behave in some way. I think this is magical thinking.  I take math to be descriptive.  A language. It works well because we created it for this very purpose (e.g. calculus). And I think its still only an approximate description at that.

Wednesday, September 17, 2014

On the nature of thought

Thoughts are causal of course.  Ultimately they can make muscles or servo motors do work in the world.  My theory of thought (see my web site www.robert-w-jones.com, cognitive scientist, theory of thought and mind) is an attempt to describe the details of this activity; causal changes to some type of memory/recording, physically comparing the contents of  buffers, creating and storing new memory patterns, etc.

Things that exist, what is real, changes to reality

Some things are defined by a list of (measurable) properties/attributes.  Some things are defined by their function/use.  A "chair" for example.  (A rock or a pile of hay might be used as a chair.)
What a thing is may well change over time.  From measuring things like size and shape and mass/weight we have moved on to measuring electric charge and then quantum mechanical "spin." We have new measurements with which we can describe an object. We also find new functions for things.  Water or dirt can now be used to shield us from nuclear radiation.  A computing device that once crunched numbers may now be used to manipulate symbols and conduct social discourse.
If a mind is given greater memory and more processing speed it will likely form more and different categories with which to describe its experiences and its world. It may abstract and compress and generalize less as it  may have less need to do so. In all these ways what is real changes.

Values

(most) animals move.
But which way should they move?
Choosing is the beginning of values and thought/intelligence.
Moving toward something sensed.  a primitive "drive"
Moving away from something sensed. a primitive "aversion"

Tuesday, September 16, 2014

The responsibility of congress

According to the U.S. constitution it is congress' job to make war, not the president's.  This is just another responsibility the republicans are shirking.

Friday, September 12, 2014

Asa H preprocessors

I have used a number of preprocessors with Asa H.  I am currently trying various (data) compressors in this role.

Thursday, September 11, 2014

Health risks

Boxing and (american) football are probably too dangerous.  They probably should be banned.

Asa H as cognitive science

"Recent developments in neuroscientific theory have suggested that cognition is inherently memory-based, where memory is fundamentally associative." (Baxter and Browne, Memory as the substrate of cognition, in Proc. of 10th Inter. Conf. on Epigenetic robotics, 2010) Asa H is just such an associative, memory-based system.

Friday, September 5, 2014

Automatic keyword discovery

Some word may occur N1 times in an entire library of documents.  The same word may appear N2 times in some single document.  Words that have the highest values of the ratio N2/N1 for a given document are the best keywords for that document.  (A stoplist may be used as a preprocessor.) Such a scheme might also be useful for routing input to a collection of specialist AIs.(22 Aug. 2014 blog)  Patterns could be counted as well as traditional words.

Tuesday, September 2, 2014

Sensor upgrades and advanced concept formation

We have taught Asa H some few hundred basic concepts (vocabulary) (see blogs of 14 Feb., 16 Feb., 12 March, 1 April, and 24 May 2013) using the simplest possible sensors (mostly Lego NXT).  We will have to upgrade (see blog of 17 July 2014) in order to learn more complex concepts (like "life", "human", etc.).

Monday, September 1, 2014

Simple cut and paste modular programming

I frequently cut and paste code snippets and modules into new programs I'm writing.  I keep an "excluded variables" list with the code library that these snippets and modules come from.  When new code is added to the library it can not contain variable names or any line numbers that have been used in other library code. (Alternatively, instead of an "excluded variables list," you can search the code library for any variable name or line number that you are wanting to use.) Gluing together modules while writing a new program consists mostly of adding lines of code that equate the (local) variables between modules.