如何在多层感知器（MLP）中反向传播？

问：

这是我的反向传播函数代码

back(target) {
// Learning rate
const learningRate = 0.1;

let delta = null;

for (let layer = this.layers.length - 1; layer >= 0; layer--) {
  const neurons = this.layers[layer];

  // Calculating gradient for output neurons is different from hidden layers
  if (layer == this.layers.length - 1) {
    delta = neurons.map((n, i) => (n - target[i]) * this.sigmoid_deriv(n));
    // Update biases of the output neurons

    this.biases[layer - 1] = this.biases[layer - 1].map(
      (b, i) => b - learningRate * delta[i]
    );

    continue;
  }

  const weights = this.weights[layer];

  const w_gradient = neurons.map((n) => delta.map((d) => n * d));

  this.weights[layer] = w_gradient.map((gradients, n) =>
    gradients.map((gradient, w) => weights[n][w] - learningRate * gradient)
  );

  // update delta for next itereation
  delta = multiply(
    dot(delta, transpose(weights)),
    neurons.map((n) => this.sigmoid_deriv(n))
  );

   if (layer > 0) {
    this.biases[layer - 1] = this.biases[layer - 1].map(
      (b, i) => b - learningRate * delta[i]
    );
   }
  }
}

我正在生成 0-1 之间的随机权重和偏差。 一切似乎都很好，但根本没有学习。 我不确定我是否正确理解了此时梯度的计算。任何帮助或指导点都是值得赞赏的。