def backward(reversedWeights: List[Tensor], predictions: List[Tensor], deltas: List[Tensor]): List[Tensor] = {
   // the first weighted layer was not influenced by any layer before, so we stop here
   if (reversedWeights.length <= 1) deltas
   else {
     // calculate an error
     val error = deltas.head.dot(reversedWeights.head.T)
     // calculate an error impact
     val delta = error * derivate(predictions.head)
     // go for next calculation and prepend a current calculation to a result list
     backward(reversedWeights.tail, predictions.tail, delta :: deltas)
  }
}

val deltas = backward(weights.reverse, prev_predictions, List(final_delta))