Inherent Limitations to LLMs (and how knowledge graphs can solve them)

[Michael DeBellis]

Someone  in the meeting today mentioned that one can ask LLMs to provide a reference and they will provide a good one. I just wanted to point out a few things based on my understanding of LLMs:

When you ask an LLM for a reference that supports a previous answer, chances are very high that the answer was not actually based on the document that the LLM provides for a reference. I've been working with a dental expert and we've done several examples where she asks questions requiring deep knowledge of dental direct restoration material products. In most cases the answers from ChatGPT are wrong and the references often seem on the surface to be relevant but  in reality when an expert carefully reads them they often are not. 

An LLM is trained on huge amounts of general knowledge that includes abstracts of many papers. But that information is "baked in" to the structure of the LLM.  There is no way to trace from any specific answer to the specific corpus documents that provided the training for that answer. 

This is also why the idea of "updating" an LLM is kind of an oxymoron. You really can't. You can feed it more data so that it gives better answers on a specific topic but you can't undo the learning it already has because again it's just baked into the structure of the network. The only way to really retrain it is to literally retrain it, i.e., rebuild it from scratch which of course is a huge undertaking. 

That is why the Retrieval Augmented Generation (RAG) architecture where you mix knowledge graphs and LLMs is so appealing. With RAG you use the LLM's capability to parse and generate natural language. But you don't use the domain knowledge of the LLM. Instead you provide a corpus of documents indexed in a database or knowledge graph. With this approach all the problems: hallucinations, black box reasoning, and bias go away. 

You get rid of hallucinations because you no longer rely on the general knowledge of the LLM which is impossible to evaluate. Instead you compute the semantic distance between the vector that models the meaning of a question with the vectors that model the meaning of the documents in your corpus. If you can't find vectors above a certain threshold you return a predefined answer that says not enough info is available. No more hallucinations. 

You get rid of black box reasoning because when you use the semantic distance metric to compute the nearest neighbors to the question vector you get specific parts of documents that are then used by the LLM to generate your answer. You can return those documents with the specific generated natural language text. 

You get rid of bias because you now have a curated set of documents that are the source of your knowledge and you can curate those documents to weed out bias. 

You can also REALLY retrain a RAG system. If you find an answer that is incorrect you can easily find the documents used to generate that incorrect answer and you can remove or rewrite them to correct the error. 

Of course the price you pay is that a RAG system is highly domain specific. Our system can answer questions about dental products but if you ask it other questions such as how to bake a cake or what Diabetes is it will just return the canned answer that it can't answer that question. Most RAG systems use a relational database to store the corpus for the RAG domain model but we've found many advantages to using a knowledge graph. E.g., we can return other objects that are relevant and present them in a graph which the user can then explore to find more information. 

Michael

[Chris Mungall]

Great summary, Michael.

I think it's a bit strong to say that RAG "gets rid of" hallucinations, but it drastically reduces opportunities for hallucination and generally decreases unreliability. However, incompleteness is still a challenge, since cosine distance over embeddings can't capture the reasoning you often need for complex queries.

RAG doesn't have to be domain-specific. In our curate-gpt framework we have a pattern where we use an LLM to generate a query over a large generic corpus such as pubmed, and then re-rank the results using RAG, essentially making a RAG database on the fly. We use this in our DRAGON-AI ontology generation system.

I am interested in catalyzing a global federated embeddings database system. Imagine that rather than making your own bespoke RAG database using llamaindex/langchain you could query a federated set of indexes, meta-rank the collected results and then do the final prompt engineering over the results. In the NCATS Biomedical Translator project we have come up with a proposed standard Vector Embedding Named Object Model indeX (venomx) to facilitate sharing embeddings in a reusable way. We're going to start publishing these for common biomedical ontologies and knowledge graphs - follow the repo if interested!

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/76123fe8-c32b-4da3-b5ad-d0d801e845f9n%40googlegroups.com.

[Michael DeBellis]

Chris, I know it is a bit strong to say RAG gets rid of hallucinations but I think I might still say it. What I mean is that hallucinations are "baked in" to the standard LLM architecture whereas in RAG you may still get wrong answers but the process isn't as opaque as with a standard LLM. I.e., a system giving you a wrong answer isn't the same thing as an hallucination. An hallucination is especially bad because except for the knowledge of the user there are never any clues that it IS an error, the LLM states it with the same confidence as any other answer. Whereas with RAG you have references as well and you can check the distance metric. E.g., if you see documents used as justification for an answer that you know are out of date you have evidence the answer may be wrong. Or if you have to set your confidence score low, say to .5 and the best answer has .5001 confidence again that's a clue. I know I'm splitting hairs here but I think it is still an important point. With RAG you don't have the same kinds of errors, but yes of course I agree errors are still possible. 

Also, good point about RAG not being just for domain specific problems. Actually that was something that surprised me when I read up on it. The first implementations of RAG systems were still for general domains they were tuned to answer specific types of general questions. E.g., I think one of the first implementations of RAG was Lewis 2020 where they still used a general domain but used RAG for four specific types of questions such as abstractive question answering and fact verification. 

I am interested in catalyzing a global federated embeddings database system.

Interesting. Another interesting development I've been reading about is using LLMs and other algorithms (e.g. A*) to build Agents and do general planning. Andrew Ng has written about this recently: https://www.deeplearning.ai/the-batch/how-agents-can-improve-llm-performance/ 

I'm attaching a paper I'm currently reading that is fascinating. They start by using the same model for agents, events, and states that I've used in previous ontologies such as Universal Moral Grammar and Cognitive Modules. I've been thinking of using Prolog to construct plans but I think the ideas in the attached paper may be more promising in terms of scaling up. 

Michael

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., . . . Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Proceedings of the 34th International Conference on Neural Information Processing Systems (pp. 9459–9474). 

You received this message because you are subscribed to a topic in the Google Groups "ontolog-forum" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/ontolog-forum/jViEL6mijv8/unsubscribe.
To unsubscribe from this group and all its topics, send an email to ontolog-foru...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAN9AifsoPpGtrX20HyjaUY1GEjCkxpJumL6yFVKUxhV_md1%3DPA%40mail.gmail.com.

[ch...@codil.co.uk]

I am currently looking at the relationship between my research into the evolutionary relationship between  CODIL and natural language and it suggests a fundamental underlying cause of the limitations of large language models.

Information is stored in the brain as links between nodes in a neural network and is exchanged between people using natural language. This means reformatting the information from a recursive network into a linear string of symbols (i.e words).  From a initial protolanguage syntax has evolved to make the linear string of symbols as compact as possible - and this compaction assumes that both the human speaker and the human hearer have a similar understanding of the context surrounding the information.

But the large language models don’t have the background context needed to convert the sequential string of symbols back into a network which accurately matched the network of the sender.
Of course if you take a vast number of word strings there are bound to be built in patterns and I feel that what the LLM are doing is, remarkably accurately, reflecting the syntax  of the messages, rather than the "understanding" contextual framework.

To put it bluntly I think that large language models have been designed to model the often complex syntax structures that have evolved to compact the linear symbolic messages between networks rather than accurately modelling the information stored in the brains of the humans that generated the messages. If this is the case, hallucinations, misunderstandings and other errors are innate in the large language model approach.

Surely if all you want is a computer system which produced good syntactically acceptable text LLM is the way to go. If you want a system that accurately handles knowledge perhaps a different approach to LLM and big data is needed - perhaps even another AI Winter and Spring.

For your information CODIL is a computer language designed to represent the network model in the human brain so information can be transferred accurately to the neural network in a friendly computer network without the need to code and then decode the information via the complex syntax structures of natural language.  For more information see https://codil-language.blogspot.com 

Chris Reynolds

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikcqC9UkJk4D20%3D2iMr4j96H5hc%3DaHOrJeim8-vYbObnhA%40mail.gmail.com.

[Michael DeBellis]

Well, I'm always the curmudgeon but as much as I know this goes against what most people accept as obviously true in cognitive science these days I don't think it is as obvious as people think:

Information is stored in the brain as links between nodes in a neural network

 Of course some information is stored in neural networks. E.g., like other primates we have face recognizers in our visual cortex and when you learn to play the piano or play tennis (i.e., muscle memory) that information is stored in a neural network. The issue I have a problem with is episodic memory like knowing which bands Eric Clapton played with or if you are a scrub jay where you buried certain kinds of food (research shows that scrub jays have an amazing memory for remembering not just where but what kind of food was stored for the winter, they dig up the more perishable foods first). 

Does anyone know of a computer neural network that can store this kind of information? I certainly don't. We store that kind of information in databases, files, etc. You might claim it is embedded in neural networks because for example they can answer questions about books like Harry Potter. But my point is that having the behavioral capability to spit out passages from Harry Potter is not the same as having a memory where you can just go and rapidly retrieve the text of the book (or the memory of reading it). 

Also, the speed of computer neurons is exponentially faster than human neurons and the amount of training given to a neural network would take an animal several lifetimes to acquire. There have been attempts to model this kind of information as neural networks. Gallistel and King examine them in their book Memory and the Computational Brain and demonstrate fundamental problems with them. The G&T analysis demonstrates that neural nets couldn't even store the amount of episodic memory that a scrub jay must have let alone a human. What's more human language and memory have the characteristic of discrete infinity. I.e., iit uses what Peirce called symbolic language which can represent an infinite number of thoughts. It is called discrete infinity because of course in practice it is still bounded by memory limitation. The kinds of codes I'm talking about are human language, DNA, and RNA. A neural net is still a finite state machine and doesn't have that capability. (Of course if you give it enough training it can seem infinite but that kind of training isn't practical for animals including humans) Some researchers such as G&T postulate that there must be some as yet undiscovered mechanism for human episodic memory that is similar (or perhaps utilizes) a biomolecular model similar to DNA and RNA. Gallistel did some fascinating research after the book where he identified some possible mechanisms in the brain and potential biomolecular codes. 

About a decade ago I audited a fascinating class called Introduction to Cognitive Neuroscience by Jack Gallant. He has his own lab at Berkeley and has done amazing work such as being able to tell what kind of picture someone is looking at  by analyzing their fMRI patterns. I asked him about this and he agreed that some form of biomolecular memory was the most likely hypothesis for episodic memory. 

Michael

https://www.michaeldebellis.com/blog 

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1715271349.698tx546ysgksk4o%40webmail.easily.uk.

[Dan Brickley]

You might dig around the literature on hippocampus eg place/grid cells, eg 

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4315928/ 

http://www.scholarpedia.org/article/Models_of_hippocampus

https://www.cell.com/neuron/fulltext/S0896-6273(00)80773-4

.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikefgN2Gve%3DSkOKnWk17HWKvs9VeHk3xQFweaUSPuZwOvg%40mail.gmail.com.

[Alex Shkotin]

Chris,

If we discuss image, audio, video processing by genAI we should not stop on LLM. We should discuss possibilities of multimodal AI where one of the last achievements is a clip generation by text prompt.

see here about Sora. And there are many other multimodal genAI. 

For GenAI the question is what is a training set? LLM is a text input, text output system.

The number of input tokens right now is from 10 to 100 thousands. To process our eye level of a picture they need 10-100 times more. And that is what genAI optimists are talking about: 10-100 times larger genAI ANN would be "a computer system which produced good syntactically acceptable" pictures, audios and videos.

About CODIL, as for any formal language, may I ask you:

-where is a Reference manual?

--especially Context-Free grammar. 

-Where is a Compiler or Interpreter to work with texts in CODIL?

-Is there a comparison with RDF?

To take language in practice we need just two things in minimum: 

-Reference Manual, where context-free grammar is inevitable;

-Compiler or Interpreter.

Alex

чт, 9 мая 2024 г. в 19:15, <ch...@codil.co.uk>:

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1715271349.698tx546ysgksk4o%40webmail.easily.uk.

[ch...@codil.co.uk]

Hi Michael

Your response to my email raises a number of important points/potential misunderstandings.

Surely the distinction between semantic and episodic memory depends on the observer.  I have no interest in, or significant knowledge of, Eric Clapton, the bands he played in, or the music that they played  (in part for reasons mentioned below). To me a data base of bands, artists and tunes is surely semantic memory rather than episodic memory - but to you thinking of Eric Clapton and his music may well invoke episodic memory of events you have attended and music you have enjoyed.

Are you suggesting that there is a rigid and important distinction between semantic and episodic memory that is independent of observer and that in order to be defined to be intelligent a system MUST be good at handling episodic information? Can you quote a scientific experiment that assesses the intelligence of two similar systems, which differ only in their ability to  handle episodic information. Without that evidence the idea that an intelligent system must be able to support episodic information is only an unproved hypothesis - and it is irrelevant that many psychologists and AI researchers believe it without proof. Could the idea that one needs to make a rigid  distinction between semantic and episodic information the current AI paradigm which will lead to another AI winter???

A  lot of what we know about the brain has been gathered by observing individuals whose brain is "different" because of illness, accident or genetic variations and there are individuals who are comparatively weak at handling episodic information and who you presumably believe lack intelligence. 

I am one of those individuals in that I have aphantasia (like an estimated 2% of the population) and cannot directly recall episodic information linked to vision, speech, smell or touch. So if the ability to handle episodic information is, as you suggest, essential to intelligence I must be exceptionally stupid. You are ignoring the fact that current research into the history of science suggests that many of the people we now consider very intelligent, such as Newton, Einstein, and possibly Darwin, were creative because they were neurodiverse.

Of course the brain is flexible and most people who have aphantasia are unaware that their brain handles episodic information in a different way to the majority. I only discovered at the age of 85 that I had aphantasia. If someone said to me that they could see something in their mind's eye I always thought they were talking metaphorically and didn’t realise that they could actually recall mental images.  It is the same for music in that I can't replay familiar tunes in my head but can recognise when a wrong note is played.

As the condition aphantasis was  only recognised recently there is much incomplete research in progress. As I see it my forebrain remembers events in terms of semantic symbols and I use this symbolic information to recall visual information as word pictures rather than images.

This is very relevant to my research into CODIL, and my current thinking is that what I have done was to model how a human brain with aphantasia processes semantic information in an intelligent way. However my current reassessment of the original research suggests that the "mental blockage" that "excludes" sensory episodic information is equivalent to a dysfunctional recursive route through the network in my brain (and in the CODIL software). 

I am currently drafting a paper on my latest ideas on CODIL, and trying to get the latest version of COGIL working again. The only reason a copy of the manual is not yet on the CODIL-language blog is that the master copy is on a floppy disc and I currently don't have acess to a computer which can read it - but I am trying to arrange access so I can post it online. I will be posting further on this issue,

Chris Reynolds

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikefgN2Gve%3DSkOKnWk17HWKvs9VeHk3xQFweaUSPuZwOvg%40mail.gmail.com.

[ch...@codil.co.uk]

Hi Alex

 

I am currently drafting a paper on CODIL which I hope to finish in about a week. In addition this email addresses  the availability of a working CODIL system and manual  and also includes a few notes I have made comparing CODIL with RDF.

 

Working CODIL System

 

The original CODIL interpreters were written for mainframe computers which no longer exist. However I got frustrated in about 1980 with AI oriented papers which described the system working and which were being rejected with comments by anonymous referees as "too theoretical ever to work" and in effect branding me as a liar for falsely claiming that the system actually worked. As the basic CODIL algorithms are very simple (effectively a search routine which crawls through a network deciding which nodes should be activated)  I decided to produce a portable version. MicroCODIL took the form of an educational package, which could be put in an envelope and sent to anyone interested - so that they could see for themselves that it really did work. MicroCODIL was trial marketed and attracted very favourable public reviews by reviewers who were not afraid to give their names!!!! 

 

In theory MicroCODIL is still available and runs on a BBC Micro - which was a very popular system at the time - and second hand computers are still widely available on ebay - although often with no monitors or disc drives and many have unofficial modifications. My problem is that while I have three BBC computers (two with faults)  I do not have a suitable working monitor or a printer. However MicroCODIL still exists on floppy discs and was still working until my monitor failed about a year ago.

 

Just over 2 years ago the plan was that I would visit an established computer archive site which has working BBC computers> The idea was that they would have security copies of the working software and manual in their archives, the fault in one of my BBC computers would be fixed, and there would be usable listable copies of the software and manual in a widely usable digital format (perhaps as pdf files). This would involved me taking my faulty BBC and a pile of floppy discs to the archive, and about one day's work to ensure that I could run CODIL and II had versions of the software and manual on my PC (making distribution easy).Jn addition the archive would hold a secure permanent record copy of the software so anyfuture researchers could have access if they needed it.. Because of the Covid restrictions the trip to the archive was cancelled and when the travel restrictions were lifted my archive contract had changed job and the help was no longer available.

 

Because of my age  (eyes less sharp and arthritic fingers) I would probably do damage if I tried to repair my faulty equipment in the very small back bedroom that I use as an office. However I am looking into the possibility of finding another archive organisation which is interested in accepting, and archiving, a permanent copy of the program, and to copy the program (and manual) from a BBC formatted floppy disc into a more widely usable format, such as a pdf file. 

 

Basically if the pre-Covid arrangements had not fallen through there would be now be no difficulty in making working copies of the MicroCODIL software and a digital copy of the manual widely available. However there are a few printed manuals left over from the trial marketing and I could let you have one - the only problem is that the manual is designed to be used by someone who has a copy of MicroCODIL and you would need to acquire a second-hand BBC Micro to run the software.  

 

Let me make it clear that I am feeling very frustrated by the current situation and the fact that I am having difficulty in providing the information you quite reasonably want. If I was 50 years younger I am sure I would have been able to move faster to sort out the problem.

 

RDF

I have drawn up some notes which show how the opening examples in the RDF Primer might be handled in CODIL The relevant secction of the primer reads:

An RDF statement expresses a relationship between two resources. The subject and the object represent the two resources being related; the predicaterepresents the nature of their relationship. The relationship is phrased in a directional way (from subject to object) and is called in RDF a property. Because RDF statements consist of three elements they are called triples.

<Bob> <is a> <person>.
<Bob> <is a friend of> <Alice>.
<Bob> <is born on> <the 4th of July 1990>. 
<Bob> <is interested in> <the Mona Lisa>.
<the Mona Lisa> <was created by> <Leonardo da Vinci>.
<the video 'La Joconde à Washington'> <is about> <the Mona Lisa>

 

CODIL items resemble RDF triples.

Each item consists of a set name,  a partition function or a "demon" describing a property of the set, and the identity of an object (where the object may be the name of another set.)

 

So we have

PERSON = Bob	Where the = is a "demon" to indicate that Bob is a member of the set PERSON.
PERSON = Alice, FRIEND = Bob	This pair of items represents that Bob is a friend of Alice
FRIEND (ISA)= PERSON	The (ISA) indicates that all members of the set FRIEND are also members of the set PERSON
FRIEND = Alice. FRIEND = Bob	Bearing in mind the previous (ISA) this means that the PERSON = Alice has a FRIEND = Bob and vice versa - i.e. that they are mutual friends.
PERSON = Bob, BIRTH DATE = 4th July 1990	The majority of CODIL statements take this form and may involve a variable number of items.
PERSON = Bob, PICTURE = Mona Lisa	This simply indicates that there is an link between the items.
VIEWER = Bob, PICTURE = Mona Lisa	In this statement Bob is placed in a set appropriate for people looking at pictures
ARTIST = Leonardo da Vinci, PICTURE = Mona Lisa
VIDEO = La joconda a Washington, PICTURE = Mona Lisa

 

The RDF primer goes on to look at how various software systems can use the RDF information. In CODIL, considered as a map of a network, where each and every node represents a set or a partition of a set. There is no conventional division into "program" and "data" as every node can act as a "data item", a "conditional test" or as a command depending on context. This allows a  CODIL network to process CODIL statements and there is no need for a separate conventional procedural language to search, amend or update CODIL files.

 

To put CODIL in context It is useful to look at the relevant dates.

 

The CODIL project was started in 1967 and was designed to provide a transparent, self documenting and human friendly interface in large commercial data processing systems but was interrupted by a major company merger, and continued on a smaller scale. In 1980 it was involved in a British Library research project called BLEND (not vto be confused with a later project with the same naame). BLEND was designed to explore the possibility of using interactive systems to write, publish and access scientific papers. CODIL was used to write the only fully interactive paper in the project. This allowed the reader of the paper to interactively run the software the paper described, and interrogate the relevant "program" and "data" files. As far as I know this was the first system where this could be done.

 

The CODIL project was closed down in 1988 because it was deemed incompatible with the way mainstream AI research was going, and the world wide web only started in March 1989. The earliest specification for RDF on the World Wide Web was in 1997 and the most recent specification was published in 2014.

 

Even if it is deemed that a study of CODIL is not relevant to solving the current problems of black box AI systems, and does not provided a simple but evolutionary compatible model for the evolution of human intelligence, the preservation of the archives could provide detailed information on how interesting creative unconventional research was lost in the politics of AI development which involved frequent changes in favoured paradigm.

Chris Reynolds

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAFxxROS8VQ16b%3D0PEYy1j6%2Bd_%2Bo93NpXFS718BJLqhhuOu_sgQ%40mail.gmail.com.

[Dan Brickley]

I do hope you can get this software archived! Are multiple copies of the floppy disks available?

In looking around for copies i found an old advert, see page 51 of 

https://ia803001.us.archive.org/2/items/AB_Computing_1988-04_OCR/AB_Computing_1988-04_OCR.pdf

Dan

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1715418784.p9pexuu7fog4gcg8%40webmail.easily.uk.

[Alex Shkotin]

Exactly!

сб, 11 мая 2024 г. в 19:43, Dan Brickley <dan...@danbri.org>:

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAFfrAFqU4RN%3DfYusVt0dtx796kHJ7WvQaLmhutpkmDXkNmEArg%40mail.gmail.com.

[Alex Shkotin]

Chris,

Dan found a nice advert [1].

For me there are two projects here.

First, to port your MicroCODIL system from your diskets to one or another modern OS. 

I think the way to begin discussion may be here http://www.bbcmicro.com/

First subtask may be to find somewhere IT configuration where the drive for your diskets is connected to a computer with modern OS.

What about this http://abug.org.uk/index.php/2020/05/16/recreated-acorn-system-3-andy-nightingale/?

The second is the CODIL Reference Manual. It is nice to have examples but not sufficient. You know.

Grammar would be perfect.

And about RDF comparison. The question is if any RDF-statement can be converted to CODIL one and vice versa?

Actually it looks like two question ✌️

Some years ago, maybe 10, I got an email from a group at the Computer Museum for BESM-6 software, on which we deployed our great DBMS Kompas in 70-th! They asked about some nuances of behavior of BESM-6 assembler code. Unfortunately I was responsible for the IBM-360 version of our DBMS.

Alex 

[1] p.51 https://ia803001.us.archive.org/2/items/AB_Computing_1988-04_OCR/AB_Computing_1988-04_OCR.pdf 

 

сб, 11 мая 2024 г. в 18:05, <ch...@codil.co.uk>:

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1715418784.p9pexuu7fog4gcg8%40webmail.easily.uk.

[Igor Toujilov]

Chris,

If the CODIL source code is available, then I am pretty sure there are software engineers who can port it to modern systems and reanimate it as their free open source project quite quickly.

Otherwise the porting might take a long time or even be impossible.

[Dan Brickley]

Is the bbc software on the floppy disks in some kind of source format, or compiled?

Dan

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAKP%2B7LMDB%3DHdzZu--4obFf7XBWKmDm7J%3D62d79nPOPc5fy-6vA%40mail.gmail.com.

[Michael DeBellis]

Surely the distinction between semantic and episodic memory depends on the observer. 

It is a distinction that many people in Cognitive Science take as a given.  

The concept of episodic memory, according to Tulving (1983, 2002), refers to a declarative memory that contains information specific to the time and place of acquisition (what laypeople may call an autobiographical episode), as distinguished from semantic memory, which is concerned with knowledge not tied to its context of acquisition.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5060006/#:~:text=The%20concept%20of%20episodic%20memory,knowledge%20not%20tied%20to%20its 

Also, Episodic memory and autonoetic consciousness

There are countless others. 

Also, I think it is a concept that we take for granted in computer science. E.g., when you create a Retrieval Augmented Generation system you store the corpus in some kind of database typically relational but occasionally with a knowledge graph. But I've never read anyone who for a minute thinks you can or should store this kind of memory in a neural net. The same for the training sets for any ML system. Those are always stored in some kind of memory addressable format like a file or database. I can't recall a single example where this isn't the case. Why not? If this kind of human memory is also stored in neural networks than why don't we also use neural networks to store data like this? There have been some attempts but to my knowledge no one has come close to a system that could be considered a replacement for files and databases and in fact most of the attempts haven't even seriously attempted this. They've tried to model the kind of memory humans have and those are the models that Gallistel and King show simply are not viable. 

Are you suggesting that... in order to be defined to be intelligent a system MUST be good at handling episodic information?

No. Because I don't think arguing about what makes a system "intelligent" are worthwhile. I agree with Turing: 

“The original question, ‘Can machines think?’, I believe to be too meaningless to deserve discussion.” Alan Turing Computing Machines and Intelligence: https://academic.oup.com/mind/article/LIX/236/433/986238  

Can you quote a scientific experiment that assesses the intelligence of two similar systems, which differ only in their ability to  handle episodic information.

 I don't think there is any point in trying to "assess... the intelligence of two similar systems" unless you first define much more clearly what your operational definition of "intelligence" is and in any case as I mentioned, I seldom find such debates worthwhile. But as for examples, if you look at the vast majority of most ML systems they have some type of addressable memory such as files or a database as well as the neural net. LLMs are one of the exceptions but they are fairly recent. For example Watson used Wikipedia and other sources that were stored in addressable memory. 

But we were talking about humans anyway.  And I take it as a given that the way to efficiently solve some problem in software is going to differ significantly from the way humans do it. That's almost universally true in general. The way we engineer artifacts to do work and solve problems is often inspired by some biological analog, and there are often underlying principles that apply to both but I don't know of a single example where there is a direct mapping between an efficient engineered solution and the way evolution solved a problem. Because evolution is a terrible engineer. Evolution is always constrained by the existing design and by the possible  viable alternative designs that can be achieved by random mutation. Human engineered artifacts aren't. 

Michael

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1715336287.o3e4nwksgwkogsog%40webmail.easily.uk.

[Ravi Sharma]

Not following too well as I am not a neuro-psychologist!

But just citing from fMRI study that recitations are expressed as semipermanent changes in brain lobes proven by Italy-India Yajurveda studies.

Comment: what is learned semantically then unlike episodic is it memorization or there are deeper learning layers/

regards.

Thanks.

Ravi

(Dr. Ravi Sharma, Ph.D. USA)

NASA Apollo Achievement Award

Former Scientific Secretary iSRO HQ

Ontolog Board of Trustees

Particle and Space Physics

Senior Enterprise Architect

SAE Fuel Cell Standards Member

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikdnOmBsBu6tpX0%2BGggrSsOPN9xhTc_jRhHMnBZ4iFKstQ%40mail.gmail.com.

[Nadin, Mihai]

Dear and respected colleagues,

The unity of the sign—i.e. standing for something else in some respect or capacity (cf. Peirce)—and the continuum of interpretation (open ended interpretant)--implies that we overcome the arbitrary reductionist strategy of dealing separately with the syntax, with semantics, and (rarely though) pragmatics. Knowledge is always and by necessity driven by pragmatics. We know for a purpose. This understanding is essential to any attempt at ontology. Algorithmic computation, which currently dominates, is by necessity of syntactic nature. Attempts are made to reach the semantic. It is actually (such as in LLM driven models) a pseudo-semantic level—understanding based on imitation, i.e. mimetic. It would be interesting to push for the pragmatic level. This is what John Sowa suggests: talk to those who in their domain of knowledge who are pursuing goals—such as making new things, initiating new processes, opening new epistemological horizons. Light can vaporize—where some of the conversation in which we are involved started—is one example. I know of many more—in biology.

Wish you well.

 

Mihai Nadin

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAAN3-5enEviH2YQtsCcmW8r4Q18bv6%3D%2B6H4V%2BKVi-umvtprwOA%40mail.gmail.com.

[Ravi Sharma]

Mihai

Yes  I agree.

Thanks.

Ravi

(Dr. Ravi Sharma, Ph.D. USA)

NASA Apollo Achievement Award

Former Scientific Secretary iSRO HQ

Ontolog Board of Trustees

Particle and Space Physics

Senior Enterprise Architect

SAE Fuel Cell Standards Member

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/BL3PR01MB689717976A13208388EFD825DAE22%40BL3PR01MB6897.prod.exchangelabs.com.

[Michael DeBellis]

Ravi, I think the simplest example in these cases is to discuss animals rather than humans because understanding humans is so much more complex and often (as in this case) the same point can be made by discussing animals. So consider Scrub Jays. They are known to be able to memorize 10,000 different locations for food cached for the winter. What's more they know not just where each bit of food was stored but whether it is the kind of food that lasts, or is perishable (they dig up the perishable food first). So that's the kind of memory I'm talking about. Can you efficiently record that kind of memory with neural nets? In Memory and The Computational Brain Gallistel and King show that any proposed model for storing such data in neurons can't scale up to what Scrub Jays can do, let alone humans. There is an analog to this kind of memory in software. Namely things like files, databases, and ontologies. We just take it for granted but we use addressable, symbolic memory to store this kind of data, not a neural network. E.g., training sets are always stored as tables in some type of file (e.g., CSV and ASCII) not as graphs.

The hypothesis is that there must be something similar to addressable, symbolic memory in animals as well and that the likely "implementation" is not neural nets but some code similar to RNA and DNA. Genetics is fascinating from a computer science standpoint. There is a Rosetta Stone for transforming nucleotide codons to amino acids as well as codons that signal "stop" codes (the end of some sequence). What is amazing is that the same code is used for organisms from Nematodes to humans and everything in between.  See: figure 2 here: https://www.nature.com/scitable/topicpage/translation-dna-to-mrna-to-protein-393/  The hypothesis is some similar code is used for memory of events rather than musle memory or things like bug detectors in frogs or face detectors in primates (all of those things clearly are neural nets). 

We have existence proofs that DNA/RNA can store this kind of memory. Several people have done this, shown that you can save and read data such as a PDF using RNA/DNA. About 20 years ago I went to a presentation at Stanford where as we entered the hall they gave each person a very small vial of water. At the start of the presentation the researcher told us that he had given us a copy of the PDF of their paper They were able to store PDFs into and read them back using RNA. This approach never turned out to be viable for actual use but it shows that using these kinds of codes to model general memory can work. 

Michael

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAAN3-5enEviH2YQtsCcmW8r4Q18bv6%3D%2B6H4V%2BKVi-umvtprwOA%40mail.gmail.com.

[Ravi Sharma]

Micheal

Yes I see the role of DNA  and see value in your comments. What a database on animals foraging!

Thanks.

Ravi

(Dr. Ravi Sharma, Ph.D. USA)

NASA Apollo Achievement Award

Former Scientific Secretary iSRO HQ

Ontolog Board of Trustees

Particle and Space Physics

Senior Enterprise Architect

SAE Fuel Cell Standards Member

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikepJhiptPaJKhUh%2BEfCQvu2ZedBNReTXu44psq-z9gA0w%40mail.gmail.com.

[Nadin, Mihai]

Scrub-jays: 200 locations:

Clayton, Nicola; Emery, Nathan & Dickinson, Anthony (2006). "The rationality of animal memory: Complex caching strategies of western scrub jays". In Hurley, Susan & Nudds, Matthew (eds.). Rational Animals?. Oxford University Press. pp. 197–216. ISBN 0198528264.

Mihai Nadin

From: ontolo...@googlegroups.com <ontolo...@googlegroups.com> On Behalf Of Ravi Sharma
Sent: Monday, May 13, 2024 6:57 PM
To: ontolo...@googlegroups.com
Subject: Re: [ontolog-forum] Inherent Limitations to LLMs (& episodic memoery)

 

Micheal

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAAN3-5dHRAhnf0gKrn6FBW5pRw%3DnH8THVr7bfzn8bwxLeHrBHg%40mail.gmail.com.

[Michael DeBellis]

Nadin, thanks for that reference, interesting paper. In: Fig. 9.2 The architecture and contents of a jay’s declarative memory, they use the term episodic memory for the kind of memory that I've been discussing. This is another example that supports my point that this concept is common in cognitive science and biology. 

Michael

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/BL3PR01MB6897B87253CE1572A61BBD88DAE32%40BL3PR01MB6897.prod.exchangelabs.com.

[John F Sowa]

Michael,

The  memory for locations is controlled by the hippocampus.   The most detailed studies have been done with rodents and other small animals.  (For more info,  Wikipedia is far more reliable than any of those GPT based things).

In the fall, squirrels bury nuts in the ground, and their hippocampus grows noticeably larger for them to remember  where they buried the nuts.  As they dig up the nuts in the winter and early spring, their hippocampus gets smaller.

The hippocampus of London taxi drivers showed a similar growth -- because they had to memorize the names and locations of all the streets in London  --  a huge amount of memorization.

But today, the taxi drivers use satellites to send information about all the streets and directions about how to reach them.  As a result, their hippocampus does not grow larger.

And by the way, each animal neuron is a very complex system that is much more complex than the artificial neural nets (aNNs),   Although each animal neuron is slower than the so-called neurons of an aNN, it is comparable in its power and complexity to a small computer.  

Therefore, an animal brain does not resemble anything like an artificial neural net.  It is more like a network of billions of microprocessors, each with a fairly large amount of storage,  Even a fruit fly has a supercomputer in its tiny head.

John

 

---

From: "Michael DeBellis" <mdebe...@gmail.com>
Sent: 5/13/24 9:03 PM

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikfPySSukpMf_ECHiXux5SiuYM4jQxKtKaS2cKOui3g1VQ%40mail.gmail.com.

[ch...@codil.co.uk]

Hi Alex

My reply delayed as there have been serious family health difficulties this end which have left me more exhausted than usual. (The problems of being 86 years old with elderly relatives!)

I hope to be able to show my BBC system (and MicroCODIL) working in the next couple of days and fully working once I have have obtained a suitable old TV as modern TV and Computer Monitors air not compatible with the BBC output.

I am also working on a graphical description of CODIL, and hope to have that draft by the end of the week.

Chris Reynolds

On Sun, 12 May 2024 10:53:40 +0300, Alex Shkotin <alex.s...@gmail.com> wrote:
 

Exactly!

 

сб, 11 мая 2024 г. в 19:43, Dan Brickley <dan...@danbri.org>:

I do hope you can get this software archived! Are multiple copies of the floppy disks available?

 

In looking around for copies i found an old advert, see page 51 of 

https://ia803001.us.archive.org/2/items/AB_Computing_1988-04_OCR/AB_Computing_1988-04_OCR.pdf

 

Dan

On Sat, 11 May 2024 at 16:05, <ch...@codil.co.uk> wrote:

Hi Alex

 

I am currently drafting a paper on CODIL which I hope to finish in about a week. In addition this email addresses  the availability of a working CODIL system and manual  and also includes a few notes I have made comparing CODIL with RDF.

I am currently looking at the relationship between my research into the evolutionary relationship between  CODIL and natural language and it suggests a fundamental underlying cause of the limitations of large language models.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1715271349.698tx546ysgksk4o%40webmail.easily.uk.

 

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAFxxROS8VQ16b%3D0PEYy1j6%2Bd_%2Bo93NpXFS718BJLqhhuOu_sgQ%40mail.gmail.com.

 

 

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1715418784.p9pexuu7fog4gcg8%40webmail.easily.uk.

 

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAFfrAFqU4RN%3DfYusVt0dtx796kHJ7WvQaLmhutpkmDXkNmEArg%40mail.gmail.com.

 

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAFxxROTPXRtR3mFMtZ8mpA62Axk71jH%3DJ319tKi4aAXF0oKc_A%40mail.gmail.com.

[Alex Shkotin]

Hi Chris

I sent you the email directly.

Alex

вт, 14 мая 2024 г. в 16:45, <ch...@codil.co.uk>:

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1715631472.9hounv3ja8cgogco%40webmail.easily.uk.

[John F Sowa]

Michael,

That distinction is independent of any observer:  "Surely the distinction between semantic and episodic memory depends on the observer."

Short answer:   For any subject S, semantic information about S is what you put in a book about the definitions and theories about S.  Episodic information is what you put in a history  about individual things and events that involve S.

Generic examples, such as "Houses usually have kitchens, bedrooms, and bathrooms"  are semantic.  But a description of a particular house that you or anybody else happened to visit last week would be episodic.  It's irrelevant whether you visited it yourself or your neighbor told you about it.

John

 

---

From: "Michael DeBellis" <mdebe...@gmail.com>
Sent: 5/13/24 12:53 PM

To: ontolo...@googlegroups.com
Subject: Re: [ontolog-forum] Inherent Limitations to LLMs (& episodic memoery)

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CALGFikdnOmBsBu6tpX0%2BGggrSsOPN9xhTc_jRhHMnBZ4iFKstQ%40mail.gmail.com.

[Michael DeBellis]

John, I agree. 

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/43b564175ce04de2821aad6beb51e545%40bestweb.net.