Nengo

I now have the University of Waterloo's Nengo spiking neural network software package running in my computer lab.

Giving Asa H names for the concepts it learns

It is easy to associate a word/name with the lowest level concepts that Asa (or a human infant) learns, things like:

collision=(sense near, bump, accelerate, hear sound "collision")

approach=(sense far, move forward, sense near, hear sound "approach")

When running Asa H we frequently watch the signals being sent from one level of the memory hierarchy to the next (see my blog of 26 Aug. 2013) which makes it possible to supply names for the higher level concepts as they become activated.  This is not possible for humans of course. Humans probably do not share exactly the same higher level (i.e. more abstract) concepts.

Does AI really need biological plausibility?

Eliasmith has criticized Markham's recent Blue Brain paper (Cell, 163, pg 1, 2015) saying "But you can get all those results with a way less complicated model."  But if your interest is in modeling general intelligence then I fear that I might say the same thing about Eliasmith's own model of the brain, Spaun. (Science, 338, 30 Nov. 2012) The very limited short term memory that humans have is something I certainly don't want to duplicate in any AI.

Numerics of promotion and tenure

If you publish a paper with a coauthor sometimes you should receive half the credit because you may have done only half as much work. On the other hand if you publish the paper with Einstein sometimes you should receive more credit (than for a single author paper) since Einstein's name suggests the paper is more valuable. Even if the coauthor is just Joe Physicist it may be important that someone, anyone, agrees with your ideas. Each paper really must be evaluated on its own merit. No fixed weighting scale is as good.

New work on focus of attention in Asa H

Each level in the Asa H hierarchy learns a set of cases and the components/features that make up the case. Each case is a vector and the features are the vector's components. In my blog of 7 Oct. 2015 I noted that we can use things like statistical measures of independence to prune features.  We can also use standard statistical measures in order to determine the relative importance of each feature in identifying a case it is found in.  Such statistical measures can then be used as weights in the (dot product or other) similarity measure that Asa uses when it compares cases.

(In other case-based reasoning systems it is not real common to see the dot product used as the similarity measure.  I have tried other similarity measures in Asa H but coming out of a physics background I have probably shown a bias toward the dot product.  It is worth noting that Jannach et. al in Recommender Systems, Cambridge Univ. Press, 2011 on page 19 say "In item-based recommendation approaches cosine similarity is established as the standard metric as it has been shown that it produces the most accurate results.")

Administrative work

I was head of physics at ESU for 4 years (and associate chair of the division of physical sciences for a part of that).  I found that I did not like administrative work. Things like scheduling, organizing, report writing, staff meetings, personnel decisions, office politics, etc.  I found this was all just a distraction from S.T.E.M. work.

Asa H

Asa informs me that indexing (which helps to organize knowledge and speed up search) is a kind of learning. This has prompted me to order a book on automatic indexing.

Chaining and creativity in Asa H

Asa H incorporates a number of learning mechanisms including "rule" chaining.  At the lower level in the memory hierarchy Asa has chained together things like :

sense dock to right --> turn right --> sense dock forward --> move forward --> dock --> recharge --> charged --> back away

(see chapter 1 of my book Twelve Papers at www.robert-w-jones.com , book)

At higher levels in the memory hierarchy Asa chains together things like:

a production system --> a universal machine --> Turing equivalent --> intelligent

This is very similar to what occurs in the "creativity machines" that I studied some years back. (see Kans. Acad. Sci. Trans., vol. 102, pg 32, 1999)  Asa's reasoning is more sophisticated, however, in that it may involve concepts of different levels of abstraction simultaneously.                                       

Search in Asa H

Search has always been an important component of AI.  When a sufficiently large spatial/temporal pattern is presented to  A.s.a.  H.  it's hierarchical memory is searched over multiple levels of abstraction.  Any retrieved matching memory will then involve concepts defined across these various levels of the hierarchy.

Is the NOT concept innate?

The NOT operation can be made innate in Asa H, i.e., 1 - dot(In,Ini).  (See my paper in Trans. Kansas Acad. Sci., vol. 109, #3/4, pg 161, 2006.) A similar thing is the inhibitory neurotransmissions found in the human brain.

Building a fifth generation computing system

A cognitive architecture that makes natural use of parallel processing (including the web/cloud?) would be an ideal fifth generation system.  A.s.a. H. (and A.v.a.) is my candidate fifth generation architecture and has run as a parallel computer.

Attention

I am experimenting with additional ways to try to restrict Asa H's focus of attention.

1. Input of natural language along with sensory stimuli.  Possibly weighting words more.
2. Of all the N inputs at time t only accept the M largest.
3. Of all the N inputs at time t use a window/spotlight to attend to only M of them.  Do not move the spotlight if similarity match or utility are high enough.

Mind as mathematician

My AI Asa H discovers, records, and manipulates spatial and temporal patterns. It uses these patterns as a language with which to describe the world.

Mathematics is a theory (or set of theories) of patterns and is used as a language (or set of languages). So one could think of a mind as a mathematician.

Quantum mechanics as dualism: Hypermind?

In a quantum computer the processing (thinking) takes place either in computers in Everett's many worlds or else in the many dimensional Hilbert space.  (Depending upon your interpretation of quantum mechanics.) If our brains were quantum computers then there just might  be a world of mind which is distinct from the physical world that our bodies occupy. (The ordinary 4 space.) This is much like the spirit-body dualism of Descartes and others.

My own view is that thought and mind are classical phenomena like those described on my website: www.robert-w-jones.com, philosopher, theory of thought and mind.

It might be interesting to run something like my AI Asa H on a quantum computer, of course.  Might this produce a hypermind in its own universe?

Feature pruning the Asa H case-base

Each level in the Asa H hierarchical memory is composed out of concepts each one of which consists of features defined on the next lower level in the hierarchy.  (The concepts can be thought of as vectors and the features are the vector's components.)

In our publications on Asa H we have described ways in which we may prune some of these features.  Standard statistical measures of independence can also be used to prune features; mutual information measures, Fisher's discrimination index, the chi-square test of independence, etc.

Enabling conversations with Asa H

I have given my AI  Asa H a kind of minimalist set of concepts based (mostly) on the Toki Pona artificial language:

"need" or "want" is defined by low robot battery and need to recharge
"away", "long" (distance), or "far" is defined by a Lego NXT ultrasonic sensor reading which is approaching 255
"near" is defined by a small input reading from an ultrasonic sensor
"strength" or "force" or "push" is defined by input from HiTechnic force sensors
"white" is defined by an input reading from a HiTechnic color sensor approaching 17
"grasp" is defined as a gripper servo closing and feeling an object with force or touch sensors
"drop" is defined as opening a gripper that had been grasping an object
"strike", or "hit" is defined by inputs from force and Lego NXT contact sensors
"home" is defined by a robot's docking station and recharger
"say" is defined by robot sound or other signal transmission
any "time" is defined by reference to the computer's clock (or an external time reference)
"move" is defined by input from encoders in a robot's drive servos and by any HiTechnic motion sensors
"taste" is defined by inputs from Vernier pH and salinity sensors
"light" or "bright" is defined by input from a Lego NXT light sensor
"knowledge" is defined by input (and output) of a computer file/case-base
"front", "back", "left", "right", "side", "top"/"on", "bottom", "body", "head", "hand", "arm", etc. are all defined by force or touch sensors on those various sides/parts of a Lego NXT robot.
Any "location" is defined by input from a Dexter industries gps module
"hot", "cold", and "temperature" are defined by input from a Vernier temperature sensor
"end" and "stop" are learned as the cessation of  some servo actions
"black" and "dark" are defined by a low input from a HiTechnic color sensor or light sensor
"work" or "active" is defined by motor activity continuing over time
"eye" is defined by the inputs from a webcam
"leg" or "foot" are defined by signals to and from the appropriate servos
"word" or "name" are defined by the set of categories and names learned for them
"path" or "road" can be defined by a line following system
"food" can be defined by the measured amount of energy stored in the robot's batteries
"eat" is defined by sensing  battery recharging
"earth" or "ground" or "floor" can be defined by setting down force or contact sensors
"wall" is defined by lateral touch or force sensor contacts and gps readings
"see" is defined by inputs from a webcam, light, or color sensors
"red", "green", and "blue" (or other colors) are given as inputs from a HiTechnic color sensor
"hear" and "sound" are defined by inputs from a Lego NXT sound sensor
"color" is defined by an input from the color sensor which is neither too high nor too low
"wind" and "air" or "fluid" are defined by the input from a Vernier anemometer
"wait" or "stay" is defined by prolonged lack of servo operation and fixed gps reading
"bump" or "acceleration" is defined by input from a HiTechnic acceleration sensor
"rotation" is defined by input from a HiTechnic gyro sensor
"north", "south", "east", "west", and "magnetism" are defined by input from HiTechnic compasses and magnetic field sensors
"turn" is defined by input from the gyro sensor, compass, and servos
"fast" and "slow" are defined by the level of inputs from various servos
"hunger" is defined by a low battery charge measurement
"pain" and "breakage" are defined by input from fine damage detecting (breakage) wires
"mouth" can be defined by the robot's battery charging contacts
"piece" can be defined by the components of a robot ("body", "arm", "gripper"/"hand", etc.)
for a virus AI like Ava 1.0 "reproduction" can be defined by disk or file copying
"parent" can be defined as the source copy when file copying occurs
"child" can be defined as the file copy
"dead servo" can be defined by zero current and zero motion when the servo is commanded
We can also detect when certain sensors are "dead".
"dead robot" can be defined by seeing when all or many servos and sensors are dead and/or many "pain" signals are input
"sense" is defined by input from any of the robot's sensors
"surface" is defined as a "wall" or "floor"
"control" can be defined by the activation/use of a PID postprocessor
"age", the robot keeps track of how long it's been in operation
"young" or "new" can be defined as an "age" less than some given number
"old" can be defined as an "age" greater than some given number
"inside" can be defined by gps values falling within a certain range
Asa H can make use of a NOT or inverse (see my paper in Trans. Kansas Acad. Sci. 109 (3/4), pg 161, 2006)  and then "live" can be defined as NOT "dead". (You can elect to define the inverse of just a limited number of signals.)
"room" or "container" is defined by "floor" and "walls"
"hard" is defined by strong "push" and small displacement
"soft" is defined by small "push" and larger displacement
"take" is learned as a sequence "grasp"-"lift"-"move"
"tool" is learned as sequences like "push"-"object"-"push"-X
"collision" is learned as a sequence "near"-"strike"-"accelerate"
"damage" is learned by a sensor pegging and/or in terms of breakage detectors ("pain")
level of "health" is learned as a combination of level of  "damage" and "hunger"
"leave" is defined by a "near" proximity measurement followed by a "far" measurement
"approach" or "arrive" is defined by a "far" proximity measurement followed by a "near" measurement
"feel" is defined by input from any force or touch sensor
"good" and "bad" are defined by the degree of activation of the  "health"  concept.
Neural network preprocessors have been trained to identify various "letters", "numbers", "shapes",  common objects  etc. "similar" and "different" can be defined by the degree of match reported by such preprocessors.
Algorithms are available to detect "faces" and "people" in images and to count them.
"group" or "many" or "large" can be defined as when a count exceeds some given number.
"lone" is a single person or face.

In some cases we would like to have additional definitions for a given concept.

Looking back on this work I think that up until now I may have underestimated the value of embodiment.

 As they are listed above, the english words for each of these concepts can be learned by association with each of the relevant input hardware signals seen by the robot. We can then hope to converse with Asa in this elementary robot language.  This is an extension of the simple communications reported in chapter 1 of my book Twelve Papers (see my website www.robert-w-jones.com  under Book).