Inductive logic programming at 30
https://arxiv.org/abs/2102.10556

The paper contains not a single reference to autoencoders!
Still they show this example:

Fig. 1 ILP systems struggle with structured examples that
exhibit observational noise. All three examples clearly
spell the word "ILP", with some alterations: 3 noisy pixels,
shifted and elongated letters. If we would be to learn a
program that simply draws "ILP" in the middle of the picture,
without noisy pixels and elongated letters, that would
be a correct program.

I guess ILP is 30 years behind the AI boom. An early autoencoder
turned into transformer was already reported here (*):

SERIAL ORDER, Michael I. Jordan - May 1986 https://cseweb.ucsd.edu/~gary/PAPER-SUGGESTIONS/Jordan-TR-8604-OCRed.pdf

Well ILP might have its merits, maybe we should not ask
for a marriage of LLM and Prolog, but Autoencoders and ILP.
But its tricky, I am still trying to decode the da Vinci code of

things like stacked tensors, are they related to k-literal clauses?
The paper I referenced is found in this excellent video:

The Making of ChatGPT (35 Year History) https://www.youtube.com/watch?v=OFS90-FX6pg

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

Hi,

One idea I had was that autoencoders would
become kind of invisible, and work under the hood
to compress Prolog facts. Take these facts:

% standard _, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9
data(seg7, [0,0,0,0,0,0,0], [0,0,0,0,0,0,0]).
data(seg7, [1,1,1,1,1,1,0], [1,1,1,1,1,1,0]).
data(seg7, [0,1,1,0,0,0,0], [0,1,1,0,0,0,0]).
data(seg7, [1,1,0,1,1,0,1], [1,1,0,1,1,0,1]).
data(seg7, [1,1,1,1,0,0,1], [1,1,1,1,0,0,1]).
data(seg7, [0,1,1,0,0,1,1], [0,1,1,0,0,1,1]).
data(seg7, [1,0,1,1,0,1,1], [1,0,1,1,0,1,1]).
data(seg7, [1,0,1,1,1,1,1], [1,0,1,1,1,1,1]).
data(seg7, [1,1,1,0,0,0,0], [1,1,1,0,0,0,0]).
data(seg7, [1,1,1,1,1,1,1], [1,1,1,1,1,1,1]).
data(seg7, [1,1,1,1,0,1,1], [1,1,1,1,0,1,1]).
% alternatives 9, 7, 6, 1
data(seg7, [1,1,1,0,0,1,1], [1,1,1,1,0,1,1]).
data(seg7, [1,1,1,0,0,1,0], [1,1,1,0,0,0,0]).
data(seg7, [0,0,1,1,1,1,1], [1,0,1,1,1,1,1]).
data(seg7, [0,0,0,0,1,1,0], [0,1,1,0,0,0,0]). https://en.wikipedia.org/wiki/Seven-segment_display

Or more visually, 9 7 6 1 have variants trained:

:- show.
_0123456789(9)(7)(6)(1)

The auto encoder would create a latent space, an
encoder, and a decoder. And we could basically query
?- data(seg7, X, Y) with X input, and Y output,

9 7 6 1 were corrected:

:- random2.
0, 0
_01234567899761

The autoencoder might also tolerate errors in the
input that are not in the data, giving it some inferential
capability. And then choose an output again not in

the data, giving it some generative capabilities.

Bye

See also:

What is Latent Space in Deep Learning? https://www.geeksforgeeks.org/what-is-latent-space-in-deep-learning/

Mild Shock schrieb:

Inductive logic programming at 30
https://arxiv.org/abs/2102.10556

The paper contains not a single reference to autoencoders!
Still they show this example:

Fig. 1 ILP systems struggle with structured examples that
exhibit observational noise. All three examples clearly
spell the word "ILP", with some alterations: 3 noisy pixels,
shifted and elongated letters. If we would be to learn a
program that simply draws "ILP" in the middle of the picture,
without noisy pixels and elongated letters, that would
be a correct program.

I guess ILP is 30 years behind the AI boom. An early autoencoder
turned into transformer was already reported here (*):

SERIAL ORDER, Michael I. Jordan - May 1986 https://cseweb.ucsd.edu/~gary/PAPER-SUGGESTIONS/Jordan-TR-8604-OCRed.pdf

Well ILP might have its merits, maybe we should not ask
for a marriage of LLM and Prolog, but Autoencoders and ILP.
But its tricky, I am still trying to decode the da Vinci code of

things like stacked tensors, are they related to k-literal clauses?
The paper I referenced is found in this excellent video:

The Making of ChatGPT (35 Year History) https://www.youtube.com/watch?v=OFS90-FX6pg

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

Hi,

Somebody wrote:

It’s a self-supervised form of ILP.
No autoencoders anywhere at all.

And, this only proofs my point that ILP doesn’t
solve the problem to make autoencoders and transformers
available directly in Prolog. Which was the issue I posted
at the top of this thread.

Subsequently I would not look into ILP for Prolog
autoencoders and transformers is my point exactly. Because
mostlikely ILP is unaware of the concept of latent space.
Latent space has quite some advantages:

- *Dimensionality Reduction:* It captures the essential
structure of high-dimensional data in a more
compact form.

- *Synthetic Data:* Instead of modifying raw data, you can
use the latent space, to generate variations for
further learning.

- *Domain Adaptation:* Well-structured latent space can
help transfer knowledge from abundant domains to
underrepresented ones.

If you don’t mention autoencoders and transformers at
all, you are possibly also not aware of the above advantages
and other properties of autoencoders and transformers.

In ILP mostlikely the concept of latent space is dormant
or blurred, since the stance is well we invent predicates,
ergo relations. There is no attempt to break

down relations further:

https://www.v7labs.com/blog/autoencoders-guide

Basically autoencoders and transformers, by imposing some
hidden layer, are further structuring relations into an
encoder and a decoder. So a relation is seen as a join.

The H is the bottleneck on purpose:

relation(X, Y) :- encoder(X, H), decoder(H, Y).

The values of H go through the latent space which is
invented during the learning process. It is not simply
the input or output space.

This design has some very interesting repercussions.

Bye

Mild Shock schrieb:

Hi,

One idea I had was that autoencoders would
become kind of invisible, and work under the hood
to compress Prolog facts. Take these facts:

% standard _, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9
data(seg7, [0,0,0,0,0,0,0], [0,0,0,0,0,0,0]).
data(seg7, [1,1,1,1,1,1,0], [1,1,1,1,1,1,0]).
data(seg7, [0,1,1,0,0,0,0], [0,1,1,0,0,0,0]).
data(seg7, [1,1,0,1,1,0,1], [1,1,0,1,1,0,1]).
data(seg7, [1,1,1,1,0,0,1], [1,1,1,1,0,0,1]).
data(seg7, [0,1,1,0,0,1,1], [0,1,1,0,0,1,1]).
data(seg7, [1,0,1,1,0,1,1], [1,0,1,1,0,1,1]).
data(seg7, [1,0,1,1,1,1,1], [1,0,1,1,1,1,1]).
data(seg7, [1,1,1,0,0,0,0], [1,1,1,0,0,0,0]).
data(seg7, [1,1,1,1,1,1,1], [1,1,1,1,1,1,1]).
data(seg7, [1,1,1,1,0,1,1], [1,1,1,1,0,1,1]).
% alternatives 9, 7, 6, 1
data(seg7, [1,1,1,0,0,1,1], [1,1,1,1,0,1,1]).
data(seg7, [1,1,1,0,0,1,0], [1,1,1,0,0,0,0]).
data(seg7, [0,0,1,1,1,1,1], [1,0,1,1,1,1,1]).
data(seg7, [0,0,0,0,1,1,0], [0,1,1,0,0,0,0]). https://en.wikipedia.org/wiki/Seven-segment_display

Or more visually, 9 7 6 1 have variants trained:

:- show.
_0123456789(9)(7)(6)(1)

The auto encoder would create a latent space, an
encoder, and a decoder. And we could basically query
?- data(seg7, X, Y) with X input, and Y output,

9 7 6 1 were corrected:

:- random2.
0, 0
_01234567899761

The autoencoder might also tolerate errors in the
input that are not in the data, giving it some inferential
capability. And then choose an output again not in

the data, giving it some generative capabilities.

Bye

See also:

What is Latent Space in Deep Learning? https://www.geeksforgeeks.org/what-is-latent-space-in-deep-learning/

Mild Shock schrieb:

Inductive logic programming at 30
https://arxiv.org/abs/2102.10556

The paper contains not a single reference to autoencoders!
Still they show this example:

Fig. 1 ILP systems struggle with structured examples that
exhibit observational noise. All three examples clearly
spell the word "ILP", with some alterations: 3 noisy pixels,
shifted and elongated letters. If we would be to learn a
program that simply draws "ILP" in the middle of the picture,
without noisy pixels and elongated letters, that would
be a correct program.

I guess ILP is 30 years behind the AI boom. An early autoencoder
turned into transformer was already reported here (*):

SERIAL ORDER, Michael I. Jordan - May 1986
https://cseweb.ucsd.edu/~gary/PAPER-SUGGESTIONS/Jordan-TR-8604-OCRed.pdf

Well ILP might have its merits, maybe we should not ask
for a marriage of LLM and Prolog, but Autoencoders and ILP.
But its tricky, I am still trying to decode the da Vinci code of

things like stacked tensors, are they related to k-literal clauses?
The paper I referenced is found in this excellent video:

The Making of ChatGPT (35 Year History)
https://www.youtube.com/watch?v=OFS90-FX6pg

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

Hi,

I first wanted to use a working title:

"new frontiers in logic programming"

But upon reflection and because of fElon,
here another idea for a working title:

"neuro infused logic programming" (NILP)

What could it mean? Or does it have some
alternative phrasing already?

Try this paper:

Compositional Neural Logic Programming
Son N. Tran - 2021
The combination of connectionist models for low-level
information processing and logic programs for high-level
decision making can offer improvements in inference
efficiency and prediction performance https://www.ijcai.org/proceedings/2021/421

Browsing through the bibliography I find:

[Cohen et al., 2017]
Tensorlog: Deep learning meets probabilistic

[Donadello et al., 2017]
Logic tensor networks

[Larochelle and Murray, 2011]
The neural autoregressive distribution estimator

[Manhaeve et al., 2018]
Neural probabilistic logic programming

[Mirza and Osindero, 2014]
Conditional generative adversarial nets

[Odena et al., 2017]
auxiliary classifier GANs

[Pierrot et al., 2019]
compositional neural programs

[Reed and de Freitas, 2016]
Neural programmer-interpreters

[Riveret et al., 2020]
Neuro-Symbolic Probabilistic Argumentation Machines

[Serafini and d’Avila Garcez, 2016]
logic tensor networks.

[Socher et al., 2013]
neural tensor networks

[Towell and Shavlik, 1994]
Knowledge-based artificial neural networks

[Tran and d’Avila Garcez, 2018]
Deep logic networks

[Wang et al., 2019]
compositional neural information fusion

Mild Shock schrieb:

Hi,

One idea I had was that autoencoders would
become kind of invisible, and work under the hood
to compress Prolog facts. Take these facts:

% standard _, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9
data(seg7, [0,0,0,0,0,0,0], [0,0,0,0,0,0,0]).
data(seg7, [1,1,1,1,1,1,0], [1,1,1,1,1,1,0]).
data(seg7, [0,1,1,0,0,0,0], [0,1,1,0,0,0,0]).
data(seg7, [1,1,0,1,1,0,1], [1,1,0,1,1,0,1]).
data(seg7, [1,1,1,1,0,0,1], [1,1,1,1,0,0,1]).
data(seg7, [0,1,1,0,0,1,1], [0,1,1,0,0,1,1]).
data(seg7, [1,0,1,1,0,1,1], [1,0,1,1,0,1,1]).
data(seg7, [1,0,1,1,1,1,1], [1,0,1,1,1,1,1]).
data(seg7, [1,1,1,0,0,0,0], [1,1,1,0,0,0,0]).
data(seg7, [1,1,1,1,1,1,1], [1,1,1,1,1,1,1]).
data(seg7, [1,1,1,1,0,1,1], [1,1,1,1,0,1,1]).
% alternatives 9, 7, 6, 1
data(seg7, [1,1,1,0,0,1,1], [1,1,1,1,0,1,1]).
data(seg7, [1,1,1,0,0,1,0], [1,1,1,0,0,0,0]).
data(seg7, [0,0,1,1,1,1,1], [1,0,1,1,1,1,1]).
data(seg7, [0,0,0,0,1,1,0], [0,1,1,0,0,0,0]). https://en.wikipedia.org/wiki/Seven-segment_display

Or more visually, 9 7 6 1 have variants trained:

:- show.
_0123456789(9)(7)(6)(1)

The auto encoder would create a latent space, an
encoder, and a decoder. And we could basically query
?- data(seg7, X, Y) with X input, and Y output,

9 7 6 1 were corrected:

:- random2.
0, 0
_01234567899761

The autoencoder might also tolerate errors in the
input that are not in the data, giving it some inferential
capability. And then choose an output again not in

the data, giving it some generative capabilities.

Bye

See also:

What is Latent Space in Deep Learning? https://www.geeksforgeeks.org/what-is-latent-space-in-deep-learning/

Mild Shock schrieb:

Inductive logic programming at 30
https://arxiv.org/abs/2102.10556

The paper contains not a single reference to autoencoders!
Still they show this example:

Fig. 1 ILP systems struggle with structured examples that
exhibit observational noise. All three examples clearly
spell the word "ILP", with some alterations: 3 noisy pixels,
shifted and elongated letters. If we would be to learn a
program that simply draws "ILP" in the middle of the picture,
without noisy pixels and elongated letters, that would
be a correct program.

I guess ILP is 30 years behind the AI boom. An early autoencoder
turned into transformer was already reported here (*):

SERIAL ORDER, Michael I. Jordan - May 1986
https://cseweb.ucsd.edu/~gary/PAPER-SUGGESTIONS/Jordan-TR-8604-OCRed.pdf

Well ILP might have its merits, maybe we should not ask
for a marriage of LLM and Prolog, but Autoencoders and ILP.
But its tricky, I am still trying to decode the da Vinci code of

things like stacked tensors, are they related to k-literal clauses?
The paper I referenced is found in this excellent video:

The Making of ChatGPT (35 Year History)
https://www.youtube.com/watch?v=OFS90-FX6pg

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

Looks like sorting of rational trees
needs an existential type, if we go full “logical”.
If I use my old code from 2023 which computes
a finest (*), i.e. non-monster, bisimulation

pre-quotient (**) in prefix order:

factorize(T, _, T) --> {var(T)}, !.
factorize(T, C, V) --> {compound(T), member(S-V, C), S == T}, !.
factorize(T, C, V) --> {compound(T)}, !,
[V = S],
{T =.. [F|L]},
factorize_list(L, [T-V|C], R),
{S =.. [F|R]}.
factorize(T, _, T) --> [].

I see that it always generates new
intermediate variables:

?- X = f(f(X)), factorize(X, [], T, L, []), write(L-T), nl. [_8066=f(_8066)]-_8066

?- X = f(f(X)), factorize(X, [], T, L, []), write(L-T), nl. [_10984=f(_10984)]-_10984

What would be swell if it would generate an
existential quantifier, something like T^([T = f(T)]-T)
in the above case. Then using alpha conversion
different factorization runs would be equal,

when they only differ by the introduced
intermediate variables. But Prolog has no alpha
conversion, only λ-Prolog has such things.
So what can we do, how can we produce a

representation, that can be used for sorting?

(*) Why finest and not corsets? Because it uses
non-monster instructions and not monster
instructions

(**) Why only pre-quotient? Because a
XXX_with_stack algorithm does not fully
deduplicate the equations, would
probably need a XXX_with_memo algorithm.

Mild Shock schrieb:

Inductive logic programming at 30
https://arxiv.org/abs/2102.10556

The paper contains not a single reference to autoencoders!
Still they show this example:

Fig. 1 ILP systems struggle with structured examples that
exhibit observational noise. All three examples clearly
spell the word "ILP", with some alterations: 3 noisy pixels,
shifted and elongated letters. If we would be to learn a
program that simply draws "ILP" in the middle of the picture,
without noisy pixels and elongated letters, that would
be a correct program.

I guess ILP is 30 years behind the AI boom. An early autoencoder
turned into transformer was already reported here (*):

SERIAL ORDER, Michael I. Jordan - May 1986 https://cseweb.ucsd.edu/~gary/PAPER-SUGGESTIONS/Jordan-TR-8604-OCRed.pdf

Well ILP might have its merits, maybe we should not ask
for a marriage of LLM and Prolog, but Autoencoders and ILP.
But its tricky, I am still trying to decode the da Vinci code of

things like stacked tensors, are they related to k-literal clauses?
The paper I referenced is found in this excellent video:

The Making of ChatGPT (35 Year History) https://www.youtube.com/watch?v=OFS90-FX6pg

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

Hi,

That is extremly embarassing. I don’t know
what you are bragging about, when you wrote
the below. You are wrestling with a ghost!
Maybe you didn’t follow my superbe link:

seemingly interesting paper. In stead
particular, his final coa[l]gebra theorem

The link behind Hopcroft and Karp (1971) I
gave, which is a Bisimulation and Equirecursive
Equality hand-out, has a coalgebra example,
I used to derive pairs.pl from:

https://www.cs.cornell.edu/courses/cs6110/2014sp/Lectures/lec35a.pdf

Bye

Mild Shock schrieb:

Inductive logic programming at 30
https://arxiv.org/abs/2102.10556

The paper contains not a single reference to autoencoders!
Still they show this example:

Fig. 1 ILP systems struggle with structured examples that
exhibit observational noise. All three examples clearly
spell the word "ILP", with some alterations: 3 noisy pixels,
shifted and elongated letters. If we would be to learn a
program that simply draws "ILP" in the middle of the picture,
without noisy pixels and elongated letters, that would
be a correct program.

I guess ILP is 30 years behind the AI boom. An early autoencoder
turned into transformer was already reported here (*):

SERIAL ORDER, Michael I. Jordan - May 1986 https://cseweb.ucsd.edu/~gary/PAPER-SUGGESTIONS/Jordan-TR-8604-OCRed.pdf

Well ILP might have its merits, maybe we should not ask
for a marriage of LLM and Prolog, but Autoencoders and ILP.
But its tricky, I am still trying to decode the da Vinci code of

things like stacked tensors, are they related to k-literal clauses?
The paper I referenced is found in this excellent video:

The Making of ChatGPT (35 Year History) https://www.youtube.com/watch?v=OFS90-FX6pg

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)