<?php
//ignore_user_abort(TRUE);
set_time_limit(0);
ini_set("precision",8);


function get_real_value($inputs,$weights){
	$addon = 0;
	foreach($inputs as $key=>$input){
		$addon += bcmul($inputs[$key],$weights[$key],8);
	}

	return 1 / (1 + exp(-1*$addon));
}


function train(&$neruons,&$weights,$input,$wout,$lr){
	$errors = array();
	$neruons[0] = $input;
	$size = sizeof($neruons);
	play_nn($neruons,$weights,$input);
	foreach($neruons[$size-1] as $key=>$o){
		$errors[$size-1][$key] = $o*(1 - $o)*($wout[$key] - $o);
	}
	for($b = $size-2;$b > 0;$b--){
		foreach($neruons[$b] as $key1 => $o){
			$add = 0;
			foreach($neruons[$b+1] as $key2=>$blah){
				$add += bcmul($weights[$b+1][$key2][$key1],$errors[$b+1][$key2],15);
			}
			$errors[$b][$key1] = $o*(1 - $o)*($add);
		}
	}
	for($b = $size-1;$b > 0;$b--){
		foreach($errors[$b] as $key=>$error){
			foreach($weights[$b][$key] as $pkey=>$we){
				$weights[$b][$key][$pkey] = $we + ($lr * $error * $neruons[$b-1][$pkey]);
			}
		}
	}
	play_nn($neruons,$weights,$input);
}
function play_nn(&$neruons,$weights,$input){
	$neruons[0] = $input;
	$size = sizeof($neruons);
	for($b = 1;$b < $size;$b++){	
		$akeys = array_keys($neruons[$b]);
		$asize = sizeof($akeys);
		for($ac=0;$ac<$asize;$ac++){
			$neruons[$b][$akeys[$ac]] = get_real_value($neruons[$b-1],$weights[$b][$akeys[$ac]]);
		}
	}
}
function make_nn(&$neruons,&$weights,$input,$layers){
	$neruons[0] = $input;
	$size = sizeof($layers);
	for($b = 1;$b < $size;$b++){
		for($i = 0; $i < $layers[$b]; $i++){
			$neruons[$b][$i] = 1;
			for($j = 0; $j < $layers[$b-1]; $j++){
				$weights[$b][$i][$j] = rand(-5,5) / 100;
			}
		}
	}
	play_nn($neruons,$weights,$input);
}

?>


Example:


<?php


$input[] = .6;
$input[] = .2;

$layer[0] = sizeof($input);
$layer[1] = 5;
$layer[2] = 2;

//build the NN
$neruons = array();
$weights = array();
make_nn($neruons,$weights,$input,$layer);


//play a example of it, you can change your input, and $neruons returns a good copy of that.
play_nn($neruons,$weights,$input);


$output[] = .6;
$output[] = .2;

$input[] = .1;
$input[] = .9;
$lr = .5; //learning rate

for($i<=0;$i<=500;$i++){
      train($neruons,$weights,$input,$output,$lr);  //trains the network to [.1,.9] -> [.6,.2]
}

print_r($neruons);

?>