probs = softmax(predictions)
    loss = cross_entropy_loss(probs, target_index)
    sub = np.zeros(predictions.shape)
    sub[target_index] = 1
    dprediction = probs - sub