The MASM Forum

General => The Campus => Topic started by: mineiro on April 05, 2017, 07:28:35 AM

Title: Simple perceptron
Post by: mineiro on April 05, 2017, 07:28:35 AM
A simple perceptron algo example, neural network.
Title: Re: Simple perceptron
Post by: Magnum on April 05, 2017, 03:44:53 PM
Your diagrams in perceptron.txt are quite impressive considering that you only used characters that can be produced by a keyboard.

Great job !! :-)

Title: Re: Simple perceptron
Post by: jj2007 on April 05, 2017, 04:45:02 PM
"I learned this..." - what does it mean?
Title: Re: Simple perceptron
Post by: Magnum on April 05, 2017, 04:46:52 PM
What time is in Italy ?

Title: Re: Simple perceptron
Post by: Siekmanski on April 05, 2017, 05:54:08 PM
Confusing our synapses with boolean algebra I think.  :biggrin:
Title: Re: Simple perceptron
Post by: mineiro on April 05, 2017, 07:40:57 PM
hello sir jj2007;
means that algorithm have learned boolean truth table.
If you only change input table on source code and feed algo with this table, so algo can 'mimic' or learn that table and produce same results as boolean logic.

;uncomment one of tables below to algorithm learn
;--------------------------------
;       lea esi,TT_OR
;       lea esi,TT_XOR
        lea esi,TT_AND                  ;boolean truth table to learn, 2 INPUTS, 1 OUTPUT
;--------------------------------

So, if algo learned AND logic, you don't need teach that algo every time, you can save synapses. So, this sounds a bit like a mutant algo, well, the code remains the same only input data changes, and this means that algorithm can realize different actions to different inputs.

This is like a cell, if we don't have boolean instructions (logical instructions) on a pc, well, we can use this code to do this job to us.
Title: Re: Simple perceptron
Post by: hutch-- on April 05, 2017, 11:57:28 PM
I wish I had time to play with AI as its very interesting stuff, working with it in text form requires some very good parsing code and the data storage is truly massive to be useful but it is fascinating stuff.
Title: Re: Simple perceptron
Post by: mineiro on April 06, 2017, 03:39:33 AM
yes sir hutch;
The nice thing about these neural network is that code is done, so we play only with data.
I can see on future persons selling trained data instead of code.
Something like an update to this simple perceptron is just inserting other logical gates (not tested), like nand,nor,... . So, trained data (synapses) can be shared.

On example posted synapse data can be stored on just 2 bytes (1 byte to each synapse) but I understood your point of view, on complex I.A. we need a lot of input data and organize code with parser to produce something more usefull.

---edited---
I tried nor and nand truth tables and do not work, algo can't learn that. I think that I can create a better trigger (activation function) or a parser, but I  do not touch on this program again, just to be simple example.