神经元数量与Hopfield网络识别模式的能力之间是否有任何关系?
我用C#编写神经网络程序,用Hopfield网络识别模式。我的网络有64个神经元。当我为2种模式训练网络时,每件事都很好用,但是当我为更多模式训练网络时,Hopfield找不到答案!
所以,根据我的代码,我如何使用Hopfield网络来学习更多模式?
我应该对此代码进行更改吗?
我的train()
功能:
public void Train(bool[,] pattern)
{
//N is number of rows in our square matrix
//Convert input pattern to bipolar
int[,] PatternBipolar = new int[N, N];
for (int i = 0; i < N; i++)
for (int j = 0; j < N; j++)
{
if (pattern[i, j] == true)
{
PatternBipolar[i, j] = 1;
}
else
{
PatternBipolar[i, j] = -1;
}
}
//convert to row matrix
int count1 = 0;
int[] RowMatrix = new int[(int)Math.Pow(N, 2)];
for (int j = 0; j < N; j++)
for (int i = 0; i < N; i++)
{
RowMatrix[count1] = PatternBipolar[i, j];
count1++;
}
//convert to column matrix
int count2 = 0;
int[] ColMatrix = new int[(int)Math.Pow(N, 2)];
for (int j = 0; j < N; j++)
for (int i = 0; i < N; i++)
{
ColMatrix[count2] = PatternBipolar[i, j];
count2++;
}
//multiplication
int[,] MultipliedMatrix = new int[(int)Math.Pow(N, 2), (int)Math.Pow(N, 2)];
for (int i = 0; i < (int)Math.Pow(N, 2); i++)
for (int j = 0; j < (int)Math.Pow(N, 2); j++)
{
MultipliedMatrix[i, j] = ColMatrix[i] * RowMatrix[j];
}
//cells in the northwest diagonal get set to zero
for (int i = 0; i < (int)Math.Pow(N, 2); i++)
MultipliedMatrix[i, i] = 0;
// WightMatrix + MultipliedMatrix
for (int i = 0; i < (int)Math.Pow(N, 2); i++)
for (int j = 0; j < (int)Math.Pow(N, 2); j++)
{
WeightMatrix[i, j] += MultipliedMatrix[i, j];
}
还有Present()
函数(此函数用于返回给定模式的答案):
public void Present(bool[,] Pattern)
{
int[] output = new int[(int)(int)Math.Pow(N, 2)];
for (int j = 0; j < N; j++)
for (int i = 0; i < N; i++)
{
OutputShowMatrix[i, j] = 0;
}
//convert bool to binary
int[] PatternBinary = new int[(int)Math.Pow(N, 2)];
int count = 0;
for (int j = 0; j < N; j++)
for (int i = 0; i < N; i++)
{
if (Pattern[i, j] == true)
{
PatternBinary[count] = 1;
}
else
{
PatternBinary[count] = 0;
}
count++;
}
count = 0;
int activation = 0;
for (int j = 0; j < (int)Math.Pow(N, 2); j++)
{
for (int i = 0; i < (int)Math.Pow(N, 2); i++)
{
activation = activation + (PatternBinary[i] * WeightMatrix[i, j]);
}
if (activation > 0)
{
output[count] = 1;
}
else
{
output[count] = 0;
}
count++;
activation = 0;
}
count = 0;
for (int j = 0; j < N; j++)
for (int i = 0; i < N; i++)
{
OutputShowMatrix[i, j] = output[count++];
}
在下面的图像中,我训练了Hopfield的角色A和P,当输入模式像A或P时,网络以真实的方式识别它们
然后我为角色C训练它:
这是每件事都出错的地方!
现在,如果我输入类似C的模式,则会出现此问题:
如果像A一样输入模式,看看会发生什么:
如果训练更多模式,整个网格变黑!
答案 0 :(得分:1)
我发现代码中只有一个错误:您只执行一次迭代的节点值计算,而不验证值是否已收敛。我已经修好了这个方法:
public bool[,] Present(bool[,] pattern)
{
bool[,] result = new bool[N, N];
int[] activation = new int[N * N];
int count = 0;
for (int i = 0; i < N; i++)
for (int j = 0; j < N; j++)
{
activation[count++] = pattern[i, j] ? 1 : 0;
}
bool convergence = false;
while (!convergence)
{
convergence = true;
var previousActivation = (int[])activation.Clone();
for (int i = 0; i < N * N; i++)
{
activation[i] = 0;
for (int j = 0; j < N * N; j++)
{
activation[i] += (previousActivation[j] * WeightMatrix[i, j]);
}
activation[i] = activation[i] > 0 ? 1 : 0;
if (activation[i] != previousActivation[i])
{
convergence = false;
}
}
}
count = 0;
for (int i = 0; i < N; i++)
for (int j = 0; j < N; j++)
{
result[i, j] = activation[count++] == 1;
}
return result;
}
这稍微改善了结果,但是也可能需要改进以异步计算值以避免循环。
不幸的是,这仍然会引入您所描述的行为。这是由称为虚假模式的现象造成的。为了让网络学习多种模式,可以考虑使用Hebb规则进行训练。您可以阅读有关Hopfield网络here和here的虚假模式,稳定性和学习情况。