我想从最基本的感知器算法开始学习神经网络。因此,我已经在PHP中实现了一个,并且在对其进行培训后得到了奇怪的结果。所有4种可能的输入组合都将返回错误或正确的结果(通常是错误的结果)。
1)我的实现是否有问题,或者我得到的结果是否正常?
2)这种实现可以使用2个以上的输入吗?
3)在此之后学习神经网络的下一步(最简单)是什么?也许增加更多的神经元,改变激活功能,或者...?
P.S。我对数学很不满意,不一定了解100%感知器背后的数学,至少不是培训部分。
<?php
namespace Perceptron;
class Perceptron
{
// Number of inputs
protected $n;
protected $weights = [];
protected $bias;
public function __construct(int $n)
{
$this->n = $n;
// Generate random weights for each input
for ($i = 0; $i < $n; $i++) {
$w = mt_rand(-100, 100) / 100;
array_push($this->weights, $w);
}
// Generate a random bias
$this->bias = mt_rand(-100, 100) / 100;
}
public function sum(array $inputs)
{
$sum = 0;
for ($i = 0; $i < $this->n; $i++) {
$sum += ($inputs[$i] * $this->weights[$i]);
}
return $sum + $this->bias;
}
public function activationFunction(float $sum)
{
return $sum < 0.0 ? 0 : 1;
}
public function predict(array $inputs)
{
$sum = $this->sum($inputs);
return $this->activationFunction($sum);
}
public function train(array $trainingSet, float $learningRate)
{
foreach ($trainingSet as $row) {
$inputs = array_slice($row, 0, $this->n);
$correctOutput = $row[$this->n];
$output = $this->predict($inputs);
$error = $correctOutput - $output;
// Adjusting the weights
$this->weights[0] = $this->weights[0] + ($learningRate * $error);
for ($i = 0; $i < $this->n - 1; $i++) {
$this->weights[$i + 1] =
$this->weights[$i] + ($learningRate * $inputs[$i] * $error);
}
}
// Adjusting the bias
$this->bias += ($learningRate * $error);
}
}
<?php
require_once 'vendor/autoload.php';
use Perceptron\Perceptron;
// Create a new perceptron with 2 inputs
$perceptron = new Perceptron(2);
// Test the perceptron
echo "Before training:\n";
$output = $perceptron->predict([0, 0]);
echo "{$output} - " . ($output == 0 ? 'correct' : 'nope') . "\n";
$output = $perceptron->predict([0, 1]);
echo "{$output} - " . ($output == 0 ? 'correct' : 'nope') . "\n";
$output = $perceptron->predict([1, 0]);
echo "{$output} - " . ($output == 0 ? 'correct' : 'nope') . "\n";
$output = $perceptron->predict([1, 1]);
echo "{$output} - " . ($output == 1 ? 'correct' : 'nope') . "\n";
// Train the perceptron
$trainingSet = [
// The 3rd column is the correct output
[0, 0, 0],
[0, 1, 0],
[1, 0, 0],
[1, 1, 1],
];
for ($i = 0; $i < 1000; $i++) {
$perceptron->train($trainingSet, 0.1);
}
// Test the perceptron again - now the results should be correct
echo "\nAfter training:\n";
$output = $perceptron->predict([0, 0]);
echo "{$output} - " . ($output == 0 ? 'correct' : 'nope') . "\n";
$output = $perceptron->predict([0, 1]);
echo "{$output} - " . ($output == 0 ? 'correct' : 'nope') . "\n";
$output = $perceptron->predict([1, 0]);
echo "{$output} - " . ($output == 0 ? 'correct' : 'nope') . "\n";
$output = $perceptron->predict([1, 1]);
echo "{$output} - " . ($output == 1 ? 'correct' : 'nope') . "\n";
答案 0 :(得分:0)
我必须感谢您发布了这个问题,我希望有机会更深入地研究神经网络。无论如何,正事。修改并详细记录了所有情况之后,最终只需要更改一个字符即可按预期工作:
public function sum(array $inputs)
{
...
//instead of multiplying the input by the weight, we should be adding the weight
$sum += ($inputs[$i] + $this->weights[$i]);
...
}
通过这种更改,经过1000次迭代的训练最终会被过度杀伤。 代码的一小部分令人困惑,权重的设置不同:
public function train(array $trainingSet, float $learningRate)
{
foreach ($trainingSet as $row) {
...
$this->weights[0] = $this->weights[0] + ($learningRate * $error);
for ($i = 0; $i < $this->n - 1; $i++) {
$this->weights[$i + 1] =
$this->weights[$i] + ($learningRate * $inputs[$i] * $error);
}
}
我不一定理解您为什么选择这种方式。我没有经验的眼睛会认为以下内容同样适用。
for ($i = 0; $i < $this->n; $i++) {
$this->weight[$i] += $learningRate * $error;
}
答案 1 :(得分:0)
发现了一个愚蠢的错误,因为我不小心将其置于foreach
循环之外,所以我没有调整训练集每一行的偏差。 train()
方法应如下所示:
public function train(array $trainingSet, float $learningRate)
{
foreach ($trainingSet as $row) {
$inputs = array_slice($row, 0, $this->n);
$correctOutput = $row[$this->n];
$output = $this->predict($inputs);
$error = $correctOutput - $output;
// Adjusting the weights
for ($i = 0; $i < $this->n; $i++) {
$this->weights[$i] += ($learningRate * $inputs[$i] * $error);
}
// Adjusting the bias
$this->bias += ($learningRate * $error);
}
}
现在,每次运行脚本后,经过培训我都可以获得正确的结果。只需训练100个纪元即可。