A.s.a. H. Blackboards

 My earliest A.s.a. H. software was feedforward and had the output of each classifier layer feeding directly into the next higher layer.* More generally it is possible to employ (one or more) blackboards. In that case external inputs** are sent to the blackboard as are the activated outputs of all the various classifier layers. The classifier layers all watch the blackboard for input.*** This allows for feedback from "higher" layers to "lower" ones. Weights can be applied to all inputs to the blackboard in order to adjust how much feedback is present, to favor input from adjacent layers, etc.

* R. Jones, Trans. Kansas Acad. Sci., vol. 109, pg. 159, 2006 and W. Ertel, Intro. to Artificial Intelligence, Springer, 2017(2nd edition), pg. 278 and the last paragraph of page 280.

** Perhaps after preprocessing.

*** See my blog of 6 October 2024.

  

Humanoids?

The usual justification for a humanoid form factor is the idea of fitting into a world designed for humans.  But even if the world doesn't, itself, change* does that really lead to a humanoid? Surely we might  like to have added eyes in the back of our heads.** And there are plenty of times I could use a third or fourth arm and hand. And, rather than holding tools why not plug and unplug various sorts of "hands?" And legs*** are good for uneven "ground", stairs, etc. but wheels are more efficient otherwise. etc. etc. etc. the list goes on.

* And that's a big if, see, for example, my blog of 10 April 2024.

** why have "heads" for that matter?

*** two legs? four?

Full Self Driving cars

It is both funny and sad to see 15 and 30 minute videos claiming to show "full self driving.". Currently Tesla is level 2, Mercedes has a level 3, and Waymo is level 4 autonomy. Any fully self driving car needs to operate routinely day after day for YEARS without causing an accident.* Otherwise who would even insure it? Even a successful unmanned drive from coast to coast would not prove much compared to what's needed.

* as many humans do

Arduino Uno Q

 I want to tryout Arduino's new SBC. It should be able to run a light version of A.s.a. H. They were not immediately listed by Sparkfun or Adafruit when I looked. Perhaps I should wait for the 4 GB/32 GB  version. 

Vision dependence

As A.s.a. H's concept grounding* is built up (from something like Wierzbicka's** "alphabet of thought,"  through Lang's Toki Pona, and then minimal English vocabulary) I find that my use of camera(s) and neural network classifier(s)*** has grown. The extent of this depending upon the environment(s) and task(s)/specialties, of course. 

* See, for example, my blog of 4 September 2024.

** And Leibniz.

*** Marked "ANN" in the diagram of my 4 September 2024 blog.

Explainable artificial intelligence

 My XAI involves producing concept maps like that of my 4 September 2024 blog* and printing out those concepts that A.s.a. H. has active as a function of time as it runs. Of course Asa can form concepts that are not shared by humans. 

* and my Trans. Kansas Acad. Sci., vol. 120, 2017, pg. 108 and blog of 14 November 2017

AI performance improvement

 When you deploy an Asa H AI agent its performance will usually improve over time.*

The performance of a deployed LLM, on the other hand, tends to degrade over time.

* Due to learning new experiences, extrapolation, etc.