Sunday, March 22, 2020

Grading online courses

Grading is an issue for fully online courses. How do you know who really did the work? I guess you could give oral exams over Skype and demand photo ID. I hate that idea. The issues associated with oral exams are well known.

Friday, March 20, 2020

One approach to NLU with A.s.a. H.

We would like to have AIs that can read texts written in human languages and learn from them. We have given A.s.a. H. the concepts/vocabulary of the Toki Pona artificial language.* There is machine translation software that translates from English to Toki Pona and from Toki Pona to English. If this can be improved it might prove adequate for our purpose.

* See my blog of 1 Oct. 2015.

Thursday, March 19, 2020

A cost of specialization

Valuable concepts learned by one specialist agent can be (and are) passed on to the next generation of agents in that specialty. Such concepts may not be useful to agents of another specialty.* They may even be harmful.

* In terms of actual physical agents I now have 40-50 small robots like those in my blog of 8 Jan. 2018.

Error correction, forgetting, and big data

As the environment changes the concepts we use to describe it must also change. We need to forget some concepts entirely.* (Things like spirits, ghosts, slaves?) In the absence of forgetting concepts are modified by averaging over lots of additional experiences.** I.e., big data.

* See my blog of 28 Oct. 2018.
** See for example my original paper on A.s.a. H, Trans. Kan. Acad. Sci., vol. 109, No. 3/4, 2006.

Tuesday, March 17, 2020

Contemplating online labs

With countries on lockdown over the coronavirus pandemic universities are trying to go entirely online. In thinking about online labs I ask myself the question: Would you want to be operated on by a surgeon who had learned surgery online? Doesn't a science curriculum require something that remains truly "hands on?"

Friday, March 6, 2020

Avoiding big data

Humans are not expected to digest anything like the amount of data that is regularly presented to artificial neural networks. So, if A.s.a. H. is to be anywhere near as intelligent as humans are it should not need to see big data either.

Monday, March 2, 2020

Unintelligent mechanical life

It is generally believed that there was a time when there was life on earth but no intelligent life. We are currently looking for such an ecosystem on Mars or elsewhere in space. Could we have an ecosystem for mechanical life without any artificial intelligences? Could such a system then support a communist style human utopia? (See my blog of  20 January 2020.)

Thursday, February 27, 2020

Innate concepts for an AI

Stanislas Dehaene argues that humans are born with certain innate, genetically hardwired concepts and that to have human level intelligence an AI will also have to have these implanted in it.* There has been a lot of work on face recognition. I have not given A.s.a. H. such a module but certainly could do so. I have used pretrained neural networks as one sort of preprocessor for A.s.a. in order to identify things like letters and numbers. The Google AIY vision kit can recognize more than a thousand common objects. (A.s.a. has the equivalent of "place neurons" that detect gps, beacons, etc.) The AIY voice kit can recognize many common vocal commands. There is much research going on with respect to natural language understanding.

A.s.a. is hierarchical. Low level regularities are learned more quickly than higher level ones. We have also played with adjusting the learning rates differently on different levels of the concept hierarchy.** When we have done some hand coding of concepts this is equivalent to giving A.s.a. innate concepts. We have sometimes given a layer in the hierarchy a two dimensional memory to allow it to create a spatial map or 2-D vision field. A.s.a. has been given an innate sense of time via time stepping and the time dilation algorithm.

A.s.a. records, updates, and employs probabilities, are they sufficient?

A.s.a.'s hierarchically organized concepts are immediately available for reuse in new combinations. I've emphasized the importance of output/actions, prediction, and extrapolation in addition to simply passively learning sensory input patterns.

A.s.a. may be more comparable to a society of humans rather than one single person.*** Agents can specialize, helping to deal with the combinatorial explosion.**** Various agents can compete against each other in each generation. A.s.a. really can multitask even if individual humans can not.

I have been continuously working on attention mechanisms. How should error correction be propagated between layers of the concept hierarchy? What should a good object concept include? Can consolidation of learning be equated with a society training a specialist agent or is more needed?

* See, for example, How We Learn, Viking, 2020. (Something of a counter argument is in my blog of 21 February 2020.) Dehaene may equate AI to deep learning neural networks and big data, the current fad. There is, of course, a lot more to AI than that.

** And a simulated annealing process.

*** Alternatively, an A.s.a. agent might be likened to one of the specialized regions in a human brain.

**** One sort of attention mechanism.

Monday, February 24, 2020

More evidence for value pluralism

The human brain makes use of multiple neurotransmitters: acetylcholine, dopamine, serotonin. While the dopamine circuit attempts to detect "good" and "bad" or "like" and "dislike" acetylcholine signals something more like "important" versus "unimportant." Vector values again.

Friday, February 21, 2020

Innate concepts

As a result of millions of years of evolutionary history the newborn human brain appears to have innate, genetically hardwired concepts of objects, numbers, probabilities, faces, language, etc.* These are a result of adaption to the specific environments that we and our animal ancestors encountered. They may not be ideal for environments we will face in the future. They may not tell us much about Kant's "thing in itself." I can give A.s.a. H. these same concepts, but should I?** I don't want my AI to BE human. The boundaries of human intelligence are partly an accident of evolutionary history. With A.s.a. I want to expand those boundaries not retain them.

* See, for example, Stanislas Dahaene, How We Learn, Viking, 2020.
** For example, number neurons that activate when they see 1 thing, or 2 things, or 3 things...

Sunday, February 16, 2020

Another very simple specialist agent

A.s.a. H. learns that collisions are to be avoided since they may cause damage. Since clutter is seen to promote collisions A.s.a. evolves a specialist to clear clutter. The algorithm for this agent is very similar to that for a toy sumo robot except that the A.s.a. agent knows to give up and move on if the obstacle proves to be immovable.

Wednesday, February 12, 2020

Alternative flyer

A small quadcopter suspended from a balloon and with an instrumentation package suspended in turn below the drone. The assembly has slightly negative buoyancy. A tether can connect the instrumentation to a computer and trickle charge the drone’s battery at any time. This flyer maneuvers slowly which is an advantage for A.s.a.

Thursday, February 6, 2020

Flight

I am hacking a DSstyles sky walker drone in order to give the A.s.a. H. society of agents a small flying robot. This particular drone is encaged which greatly simplifies repeated takeoffs and landings. As a result of having an anemometer and microphone nearby A.s.a. immediately associates "flying" with "wind" and "engine noise" in its concept hierarchy. A.s.a. had already associated larger vertical motions with an atmospheric pressure decrease.

Saturday, February 1, 2020

Evolving robot explorers

The A.s.a. H. society of agents learns to specialize. One of the mobile robotic arms we have available has been used to transport some of the larger sensors; things like Geiger tubes, metal detectors, anemometers, etc. A.s.a. H. learns/creates a specialist “explorer agent” making use of these hardware components and uses it to probe previously unmapped areas.* The program this particular agent learns is relatively simple, mostly data logging and gps and/or beacon signal logging.

* Seeking out things like abundant light for solar panels, moderate temperatures, low clutter environment, etc. in order to maximize utility.

Tuesday, January 28, 2020

A conscious machine

Working within Baars' global workspace theory Barthelmess, Furbach, and Schon argue* that the Hyper reasoning system, with ConceptNet as its knowledge base, is conscious. While I agree with much of this I do believe there are different degrees of consciousness. I have also argued** that consciousness is a collection of processes, not one single thing. Hyper and ConceptNet does not have a notion of self*** nor does it have all 10 of Hobson's "functional components".****

I don't think that consciousness is as difficult as the "hard problem" people would have us believe. On the other hand I don't think that Hyper-ConceptNet is as fully conscious as A.s.a. H. is.*****

The attention issue is part of dealing with the curse of dimensionality. Its a problem that must be faced by any machine trying to operate in a large state space.

* arXiv:2001.09442v1, 26 Jan. 2020
** See my blog of 19 Oct. 2016
*** See Trans. Kansas Academy of Sci., 2017, page 108
**** For example, it seems to lack orientation, emotion, and values.
*****But ConceptNet is a large knowledgebase of almost 3 million axioms in first order logic!

Monday, January 20, 2020

The Communist Utopia

The argument goes something like this:
- Society requires that most of us work.
- But physics tells us that work is energy. “Labor saving appliances” allow us to replace human labor with other energy sources.
- It might be possible to make energy free. Tesla thought that there might be sources of free cosmic energy. Much of his physics was unsound but solar energy is a possible example. Lewis Strauss, the chairman of the atomic energy commission (1954), thought nuclear energy might become “too cheap to meter.” Plentiful thorium or deuterium fuels, for example.
- No one then need work any longer. Machines would replace all human labor. (Today machines are able to do half of all human jobs. But completing the task might involve the creation of  “mechanical life” and the subsequent class struggle between humans and AIs.)

Sunday, January 12, 2020

Vector values

The idea that humans have a vector value system* receives some support from Shalom H. Schwartz's "circular model of values." (see, for example, Journal of Research in Personality, June 2004, pg 230-255)

*A.s.a. H. frequently makes use of a vector value system (see my blog of 19 Feb. 2011) and my criticism of capitalism is based in part on the need to avoid a scalar utility (see my paper  www.robert-w-jones, philosopher, Capitalism is Wrong).

Friday, January 10, 2020

An example of learned attention, attending to

A.s.a. H. learns that (a robot's) collisions correlate with increased pain and damage.
It also learns that sweeping the ultrasonic (obstacle) sensor back and forth correlates with having fewer collisions as compared with having a fixed directed ultrasonic sensor. A.s.a. H. then learns to sweep it's sensor, looking for obstacles and spending more time attending to this particular input channel.

Alternatively, if the robot has a single fixed mounted sensor it may learn to make small repeated left and right turns as it advances forward.

Thursday, January 2, 2020

A kind of intentional thought

Whenever A.s.a. H. learns a case (sequence) this will include any actions that were taken.  Actions need not be the activation of servo motors, they can include things like choosing to perform “thinking with a simulation” (see my 30 May 2019 blog), or adjusting things like time spent extrapolating, doing feature extraction, etc. (e.g., adjusting parameters like L and skip, see my 10 Feb. 2011 blog) See also my book Twelve Papers, pages 15 and 16, self monitoring, www.robert-w-jones.com.

Wednesday, January 1, 2020

Disembodied AI, a complication

Following up on my 1 May 2019 blog I have replaced A.s.a. H.’s lowest layer with human inputs. Unfortunately, some common human inputs need to go to A.s.a.’s second, third, and fourth layers. This complicates learning among other things.

A.s.a. H. learns behavior trees

Colledanchise and Ogden have discussed the advantages of behavior trees in their book Behavior Trees in Robotics and AI: An Introduction (arxiv 1709.00084v3 15 Jan. 2018). Advantages are said to include modularity, hierarchical organization, reusability, and reactivity. A.s.a. H. learns behavior trees similar to that of figure 1.1 from Colledanchise and Ogden’s book:

Pick and place = (Grasp ball -> Carry ball -> Drop ball)
Grasp ball = (detect ball inside grippers -> close grippers -> sense force against grippers)
Carry ball = (sense force against grippers -> move)
Drop ball = (sense force against grippers -> open grippers -> sense no force against grippers)


A.s.a. H.'s devided self

In some experiments I have employed a society of Asa agents. In others a small processor (a LEGO NXT, EV3, Arduino, or Raspberry Pi) rode on each mobile effector (or sensor array). These little brains then were linked (frequently by a power and communication tether) to a larger processor (brain). (Somewhat like in an octopus.) The self concept that A.s.a. H. forms (see, for example, my blogs of  21 July 2016 and 1 January 2017) is then distributed among multiple brains in multiple locations.