Older blog entries for mentifex (starting at number 108)

The predictive textlike brain mechanism mentioned in the article works perhaps because each word as a concept has many associative tags over to other concept-words frequently used in connection with the triggering word. A similar neural network of associative tags is at work in the Mind.Forth Strong AI which has been ported into Strawberry Perl 5 and which you may download free-of-charge in order to study the Theory of Mind depicted in the diagram below also available as an animated brain-mind GiF:
./^^^^^^^\..SEMANTIC MEMORY../^^^^^^^\
| Visual. | .. syntax ..... |Auditory |
| Memory..| .. /..\---------|-------\ |
| Channel.| . ( .. )function|Memory | |
| . . . . | .. \__/---/ . \ | . . . | |
| . /-----|---\ |flush\___/ | . . . | |
| . | . . | . | |vector | . | .word | |
| ._|_ .. | . v_v_____. | . | .stem | |
| / . \---|--/ . . . .\-|---|--/ .\ | |
| \___/ . | .\________/ | . | .\__/ | |
| percept | . concepts _V . | .. | .| |
| . . . . | . . . . . / . \-|----' .| |
| . . . . | . . . . .( . . )| ending| |
| . . . . | inflection\___/-|->/..\_| |
| . . . . | . . . . . . . . | .\__/.. |
Syntax generates thought from concepts.
AI Mind Maintainer jobs will be like working in a nuclear power plant control room.

Yesterday we may have finally learned how to let the Ghost Perl AI think indefinitely without stopping to wait for a human user to press "Enter" after typing a message to the AI Mind. We want the Perlmind only to pause periodically in case the human attendant wishes to communicate with the AI. Even if a human types a message and fails to press the Enter-key, we want the Perl AI to register a CR (carriage-return) by default and to follow chains of thought internally, with or without outside influence from a human user.

Accordingly today we create the AudListen() module in between the auditory memory modules and the AudInput() module. We move the new input code from AudInput() into AudListen(), but the code does not accept any input, so we remove the current code and store it in an archival test-file. Then we insert some obsolete but working code into AudListen(). We start getting primitive input like we did yesterday in the ghost181.pl program. Then we start moving in required functionality from the MindForth AI, such as the ability to press the "Escape" key to stop the program.

Eventually we obtain the proper recognition and storage of input words in auditory memory, but the ghost182.pl AI is not switching over to thinking. Instead, it is trying to process more input. Probably no escape is being made from the AudInput() loop that calls the AudListen() module. We implement an escape from the AudInput() module.

The ghost182.pl program is now able take in a sentence of input and generate a sentence of output, so we will upload it to the Web. We still need to port from MindForth the code that only pauses to accept human input and then goes back to the thinking of the AI.

As an independent scholar in polyglot artificial intelligence, I have just today on March 21, 2017, stumbled upon a possible algorithm for implementing machine translation (MT) in my bilingual Perlmind and MindForth programs. My Ghost Perl AI thinks heretofore in either English or Russian, but not in both languages interchangeably. Likewise my Forth AI MindForth thinks in English, while its Teutonic version Wotan thinks in German.

Today like Archimedes crying "Eureka" in the bathtub, while showering but not displacing bath-water I realized that I could add an associative tag mtx to the flag-panel of each conceptual memory engram to link and cross-identify any concept in one language to its counterpart or same concept in another language. The mtx variable stands for "machine-translation xfer (transfer)". The AI software will use the spreading-activation SpreadAct module to transfer activation from a concept in English to the same concept in Russian or German.

Assuming that an AI Mind can think fluently in two languages, with a large vocabulary in both languages, the nub of machine translation will be the simultaneous activation of semantically the same set of concepts in both languages. Thus the consideration of an idea expressed in English will transfer the conceptual activation to a target language such as Russian. The generation modules will then generate a translation of the English idea into a Russian idea.

Inflectional endings will not pass from the source language directly to the target language, because the mtx tag identifies only the basic psi concept in both languages. The generation modules of the target language will assign the proper inflections as required by the linguistic parameters governing each sentence being translated.

We start today by taking the 336,435 bytes of ghost176.pl from 2017-03-14 and renaming it as ghost177.pl in a text editor. Then in the Windows XP MS-DOS prompt we run the agi00045.F MindForth program of 166,584 bytes from 2016-09-18 in order to see a Win32Forth window with diagnostic messages and a display of "you see dogs" as input and "I SEE NOTHING" as a default output. From a NeoCities upload directory we put the agi00045.F source code up on the screen in a text editor so that we may use the Forth code to guide us in debugging the Perl Strong AI code.

Although in our previous PMPJ entry from yesterday we recorded our steps in trying to get the Perl AudRecog mind-module to work as flawlessly as the Forth AudRecog, today we will abandon the old Perl AudRecog by changing its name and we will create a new Perl AudRecog from scratch just as we did with the Forth AudRecog in 2016 when we were unable to tweak the old Forth AudRecog into a properly working version. So we stub in a new Perl AudRecog() and we comment out the old version by dint of renaming it "OldAudRecog()". Then we run "perl ghost177.pl" and the AI still runs but it treats every word of both input and output as a new concept, because the new AudRecog is not yet recognizing any English words.

Next we start porting the actual Forth AudRecog into Perl, but we must hit three of our Perl reference books to learn how to translate the Forth code testing ASCII values into Perl. We learn about the Perl "chr" function which lets us test input characters as if they were ASCII values such as CR-13 or SPACE-32.

Now we have faithfully ported the MindForth AudRecog into Perl, but words longer than one character are not being recognized. Let us comment out AudMem() by naming it OldAudMem() and let us start a new AudMem() from scratch as a port from MindForth.

We port the AudMem code from Forth into Perl, but we may not be getting the storage of SPACE or CR carriage-return.

Now into our third day in search of stable Perlmind code, we take the 344,365 bytes of ghost177.pl from 2017-03-15 and we save a new file as the ghost178.pl AI. We will try to track passage of characters from AudInput to AudMem to AudRec.

Through diagnostic messages in AudRecog, we discovered that a line of code meant to "disallow audrec until last letter of word" was zeroing out $audrec before the transfer from the end of AudRecog to AudMem.

In a departure from MindForth, we are having the Perl AudRecog mind-module fetch only the most recent recognition of a word. In keeping with MindForth, we implement the auditory storing of a $nxt new concept in the AudInput module, where we also increment the value of $nxt instead of in the NewConcept module.

Today we return to Perl AI coding after updating the MindForth code in July and August of 2016. In Forth we re-organized the calling of the subordinate mind-modules beneath the MainLoop module so as no longer to call the Think module directly, but rather to call the FreeWill module first so that eventually the FreeWill or Volition module will call Emotion and Think and Motorium.

We have discovered, however, that the MindForth code properly handles input which encounters a bug in the Perl code, so we must first debug the Perl code. When we enter, "you see dogs", MindForth properly answers "I SEE NOTHING", which is the default output for anything involving VisRecog since we have no robot camera eye attached to the Mind program. The old Perl Mind, however, incorrectly recognizes the input of "DOGS" as if it were a form of the #830 "DO" verb, and so we must correct the Perl code by making it as good as the Forth code. So we take the 335,790 bytes of ghost175.pl from from 2016-08-07 and we rename it as ghost176.pl for fresh coding.

We start debugging the Perl AudRecog module by inserting a diagnostic message to reveal the "$audpsi" value at the end of AudRecog. We learn that "DOGS" is misrecognized as "DO" when the input length reaches two characters. We know that MindForth does not misrecognize "DOGS", so we must determine where the Perl AudRecog algorithm diverges from the Forth algorithm. We are fortunate to be coding the AI in both Forth and Perl, so that in Perl we may implement what already works in Forth.

In Perl we try commenting out some AudRecog code that checks for a $monopsi. The AI still misrecognizes "DOGS" as the verb "DO". Next we try commenting out some Perl code that declares a $psibase when incoming word-length is only two. The AI still misrecognizes. Next we try commenting out a declaration of $subpsi. We still get misrecognition. We try commenting out another $psibase. Still misrecognition. We even try commenting out a major $audrec declaration, and we still get misrecognition. When we try commenting out a $prc declaration, AudRecog stops recognizing the verb "SEE". Then from MindForth we bring in a provisional $audrec, but the verb "SEE" is not being recognized.

Although in the MS-DOS CLI prompt we can evidently not run MindForth and the Perlmind simultanously, today we learn that we can run MindForth and leave the Win32Forth window open, then go back to running the Perl AI. Thus we can compare the diagnostic messages in both Forth and Perl so as to further debug the Perl AI. We notice that the Forth AudMem module sends a diagnostic message even for the blank space ASCII 32 even after "SEE", which the Perl AI does not do.

We may need to add a tru tag to the conceptual flag-panel in the various AI Minds, such as in Forth and in Perl. Only the first word in the thought-engram will need a tru tag. We may want to have the following tags in the panel.

tru psi act hlc pos jux pre iob tkb seq num mfn dba rv

Active code will probably assign a numeric "true" value, so that only the most current thoughts will have an assumption of truth and believability. Preterite be-verb assertions like "Kilroy is here" shall have decayed down to a low tru value so that they will not be taken at face value by the thinking Mind. On the other hand, non-be-verb knowledge about the ontology of the world will need to be regarded as true.

As the thinking AI associates from thought to thought, sentence-engrams with a low truth-value should not come into play. Various criteria may cause some engrams to go to a mid-range truth-value and other engrams to a minimal truth-value, so that reliable knowledge may come into play.

The tru-tag will permit rather elaborate ideas to emerge back into consciousness with emphasis on special considerations such as the inclusion of a prepositional phrase in the idea, as in a sentence like, "A man with a boat needs money". The 3D AI will therefore need not only modules for thinking with prepositional phrases, but also modules for conjunctions to be used in sentences like, "I know that time is money" or "I think that boats cost money." The routines for comprehension will need to be modernized or adjusted to allow parts of a long input sentence to be comprehended upon selection of a likely subject.

To some extent, we are aiming for a conscious AI Mind that realizes that it lives inside a computer and that it has only limited interaction with the outside world. It may need the ASCII bell-function as a way of deliberately summoning the attention of a human user.

In agi00031.F we are trying to figure out why we have lost the functionality of ending human input with a 13=CR and still getting a recognition of the final word of the input. We compare the current AudMem code with the agi00026.F version, and there does not seem to be any difference. Therefore the problem must probably lie in the major revisions made recently to the AudInput module.

From the diagnostic report messages that appear when we run the agi00031.F, it looks as though the 13=CR carriage return is not getting through from the AudInput module to the AudMem module. When we briefly insert a revealing diagnostic into the agi00026.F AudMem start, we see from "g AudMem: pho= 71" and "o AudMem: pho= 79" and "d AudMem: pho= 68" and "AudMem: pho= 13" that the carriage-return is indeed getting through. Therefore in AudInput we need to find a way of sending the final 13=CR into AudMem. Upshot: It turns out that in AudInput we only had to restore "pho @ 31 > pho @ 13 = OR IF \ 2016aug27: CR, SPACE or alphabetic letter" as a line of code that would let 13=CR be one of the conditions required for calling the AudMem module.

Next in the InStantiate module we need to remove a test that only lets words with a positive "rv" recall-vector get instantiated, because we must set "rv" to zero for personal pronouns being re-interpreted as "you" or "I" during communication with a human user. Apparently the Perlmind just ignores the engrams with a zero "rv" and finds the correct forms with a search based on parameters.

Now we would like to see how close we are to fulfilling all the conditions for a proper "trough" of inhibition in the AI MindGrid. When we run the ghost175.pl Perl AI and we enter "You know God," we see negative activations in thepresent-most trough of both the input and the concepts of "I HELP KIDS" as the output. In the Forth AGI, we wonder why do not see any negative activations in the present-most trough. Oh, we were not yet bothering to store the "act" activation-level in the Forth InStantiate module. We insert the missing necessary code, and we begin to see the trough of inhibition in both the recent-most input and the present-most output.

Recently we have developed the ability to visualize the MindGrid as Theater of Neuronal Activations. At the most recent, advancing front of the MindGrid, we see an inhibited trough of negative activations. We see an input sentence from a human user activating concept-fibers stretching back to the earliest edge of the MindGrid. We see an old idea becoming fresh output and then being inhibited into negative activation at its origin. We see outputs of the AGI passing through ReEntry() to re-enter the Mind as inhibited engrams while re-activating old engrams. We see the front-most trough of inhibition preventing the most recent ideas from preoccupying and monopolizing the artificial consciousness.

In ghost 174.pl, we have now commented out some code in the InStantiate() mind-module that was letting only nouns or pronouns of human input be re-activated along the length of the MindGrid. The plan now is to let all parts of an incoming sentence re-activate the engrams of its component
concepts.

Now, how do we make sure that the front-most engrams of the sentence of human input will be inhibited with negative activation in the trough of recent mental activity on the MindGrid? It appears that InStantiate() makes a sweep of old engrams to set a positive activation, and then
at the $tult penultimate-time it sets an activation for the current, front-most input. In order to keep a trough of recent inhibition, let us try setting a negative activation at the $tult time-point.

After input of "I see kids" and a response by the AI of "KIDS MAKE ROBOTS", in minddata.txt we see the sweep of positive activation of old engrams.

At t=477, "YOU" has an activation of thirty (30).

At t=518, "YOU" has an activation of thirty (30).

At t=317, 820=SEE has an activation of thirty (30).

At t=575, 528=KIDS has an activation of 62, apparently because there was also a re-entry of "KIDS".

As a result of the $tult trough-inhibition,

at t=2426, 707=YOU has a negative "-46" activation.

At t=2430, 820=SEE has a negative "-46" activation.

At t=2435, 528=KIDS has a negative -14 activation, apparently because the AI response of "KIDS MAKE ROBOTS" made a backwards sweep to impose a positive thirty-two (32) points of activation upon the pre-existing negative "-46"
points of activation, resulting in -46+32 = -14 negative points of activation -- still part of the negative trough.

Now the AGI is making its series of innate self-referential statements ("I AM A PERSON"; "I AM A ROBOT"; I AM ANDRU"; I HELP KIDS") but why is it not using SpreadAct() to jump from the reentrant concept of "KIDS" to the innate idea of "KIDS MAKE ROBOTS"? Let us see if SpreadAct() is being called, and from where. We do not see SpreadAct() being called in the diagnostic messages on-screen while
we run the AGI. Let us check the Perlmind source code. We see that the OldConcept() module since ghost162.pl was calling SpreadAct() for recognized nouns, but now we delete that snippet of code because we see in our MindGrid theater that we do not want OldConcept() to make any calls to SpreadAct(). The AGI still runs.

We see that SpreadAct() is potentially being called from the ReEntry() mind-module, but the trigger is not working properly, so we change the trigger. Then we get SpreadAct() re-activating nouns, and we begin to see a periodic association from the innate self-referential statements to "KIDS MAKE ROBOTS" and from there to "ROBOTS NEED ME". Apparently the inhibitions have to be cancelled out before the old memories can re-surface in the internal chains of
thought of the AGI.

It was fun but nevertheless sincere to post AI Has Been Solved on April Fool's Day ten years ago. Mentifex Strong AI always was and always will be an extremely serious AI Lab Project as described in December of 1998 by the Association for Computing Machinery. Mentifex AI is so extremely serious that it has meanwhile been ported into Russian and into German. The resulting Amazon Kindle e-book, Artificial Intelligence in German, has been reviewed with the maximum highest-possible five-star rating. Another e-book, InFerence, describes how the Mentifex AI Minds can think by automated reasoning with logical inference. The MindForth AI prior art program has been cited in a Google patent. Now finally at http://ai.neocities.org/AiSteps.html a third-generation (3G) Mentifex AI Mind is being created in Perl, and Netizens from all over the world are looking into the use of Unicode and Perl to create artificial intelligence in any programming language and in any natural human language. Ladies and gentlemen, start your AI engines.

If your humanoid robot needs an AI Mind to think in English or German, a new Amazon Kindle e-book goes into great detail about robotic thought processes.

This e-book in English about AI in German (and English and Russian) contains the entire AI source code in Forth, which causes most of the editorial portion of the e-book (18 of 20 chapters) to be readable without charge in the free preview.