What made me do the lillte prototype? Try this one,
it has a little Java code. But its a little ancient
technologie using the sigmoid activation function. And
it seems to me it uses some graph datastructure:

Neural Networks
Rolf Pfieffer et al. - 2012

https://www.ifi.uzh.ch/dam/jcr:00000000-7f84-9c3b-ffff-fffffb34b58a/NN20120315.pdf

I guess it corresponds to this here, which is a SWI-Prolog and C
hybrid, when using FANN_SIGMOID:

FANN - Fast Artificial Neural Network
Package for SWI-Prolog - 2018
https://www.swi-prolog.org/pack/list?p=plfann

Translating the Java code to Prolog from the Pfeiffer
paper into linear algebra using vectors and matrixes, I
have now a little piece of pure Prolog code, that runs

also in the Browser, that can already learn an
AND, and its using the ReLU activation function,
i.e. not the FANN_SIGMOID activation function anymore.

I simulated the bias by an extra input neuron
which is always 1, because I was to lazy to have
bias in the model. Sample output:

A -- 0.99 ---\
\
B -- 0.99 -----+-- ReLu -->
/
1 -- -0.98 --/

It can als learn an XOR. Libraries such as PyTorch
cooperate with optimizer libraries that provide a
variety of gradient search methods. One needs

to study how these library are architectured so that
they provide plug and play. Maybe can bring the same
architecture to Prolog:

A Gentle Introduction to torch.autograd

Next, we load an optimizer, in this case SGD with a
learning rate of 0.01 and momentum of 0.9. We register all
the parameters of the model in the optimizer.

optim = torch.optim.SGD(model.parameters(), lr=1e-2, momentum=0.9)

https://pytorch.org/tutorials/beginner/blitz/autograd_tutorial.html

Mild Shock schrieb:

Ok some progress report here. I have currently a
library(linear) in the working which is only a few
lines of code, but it provides vectors and matrixes.
One can use the library to define matrix exponentiation:

matexp(M, 1, M) :- !.
matexp(M, N, R) :- N mod 2 =:= 0, !,
   I is N // 2,
   matexp(M, I, H),
   matmul(H, H, R).
matexp(M, N, R) :-
   I is N-1,
   matexp(M, I, H),
   matmul(H, M, R).

And then do fancy stuff like answering the question
what are the last 8 digits of fibonacci(1000000):

?- time((fib(1000000, _X), Y is _X mod 10^8)).
% Zeit 28 ms, GC 0 ms, Lips 88857, Uhr 16.03.2025 22:48
Y = 42546875

The 28 ms execution time are not bad, since modulo was not
integrated into matexp/3, making it to compute the full
fibonacci(1000000) before taking the modulo. Not sure whether
JavaScript bigint is faster or slower than GMP ?

So what can we do with library(linear) besides implementing
eval/3 and back/3 ? We can finally update a neural network
and do this iteratively. Using a very simple random pick
to choose some training data sample:

update([V], _, [V])  :- !.
update([V,M|L], [_,M3|R], [V,M4|S]) :-
   maplist(maplist(compose(add,mul(0.1))), M3, M, M4),
   update(L, R, S).

iter(0, _, N, N) :- !.
iter(I, Z, N, M) :-
   random(R), K is floor(R*4)+1,
   call_nth(data(Z, X, Y), K),
   eval(N, X, U),
   back(U, Y, V),
   update(U, V, W),
   J is I-1,
   iter(J, Z, W, M).

Disclaimer: This is only a proof of concept. It mostlikely
doesn’t have all the finess of Python torch.autograd. Also
it uses a very simple update of the weights via μ Δwij with
μ = 0.1. But you can already use it to learn an AND

or to learn an XOR.

Mild Shock schrieb:

Somehow I shied away from implementing call/n for
my new Prolog system. I thought my new Prolog system
has only monomorphic caches , I will never be able to

replicate what I did for my old Prolog system with
arity polymorphic caches. This changed when I had
the idea to dynamically add a cache for the duration

of a higher order loop such as maplist/n, foldl/n etc…

So this is the new implementation of maplist/3:

% maplist(+Closure, +List, -List)
maplist(C, L, R) :-
    sys_callable_cacheable(C, D),
    sys_maplist(L, D, R).

% sys_maplist(+List, +Closure, -List)
sys_maplist([], _, []).
sys_maplist([X|L], C, [Y|R]) :-
    call(C, X, Y),
    sys_maplist(L, C, R).

Its similar as the SWI-Prolog implementation in that
it reorders the arguments for better first argument
indexing. But the new thing is sys_callable_cacheable/1,

which prepares the closure to be more efficiently
called. The invocation of the closure is already
quite fast since call/3 is implemented natively,

but the cache adds an itch more speed. Here some
measurements that I did:

/* SWI-Prolog 9.3.20 */
?- findall(X,between(1,1000,X),L), time((between(1,1000,_),
    maplist(succ,L,_),fail; true)), fail.
% 2,003,000 inferences, 0.078 CPU in 0.094 seconds

/* Scryer Prolog 0.9.4-350 */
?- findall(X,between(1,1000,X),L), time((between(1,1000,_),
    maplist(succ,L,_),fail; true)), fail.
     % CPU time: 0.318s, 3_007_105 inferences

/* Dogelog Player 1.3.1 */
?- findall(X,between(1,1000,X),L), time((between(1,1000,_),
    maplist(succ,L,_),fail; true)), fail.
% Zeit 342 ms, GC 0 ms, Lips 11713646, Uhr 10.03.2025 09:18

/* realla Prolog 2.64.6-2 */
?- findall(X,between(1,1000,X),L), time((between(1,1000,_),
     maplist(succ,L,_),fail; true)), fail.
% Time elapsed 1.694s, 15004003 Inferences, 8.855 MLips

Not surprisingly SWI-Prolog is fastest. What was
a little surprise is that Scryer Prolog can do it quite
fast, possibly since they heavily use maplist/n all

over the place, they came up with things like '$fast_call'
etc.. in their call/n implementation. Trealla Prolog is
a little bit disappointing at the moment.

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