private void stochasticUpdateStep(Pair<Integer, Set<Integer>> wordPlusContexts, int s) {
      double eta = learningRateDecay(s);
      int wordIndex = wordPlusContexts.getFirst(); // actual center word
      // Set h vector equal to the kth row of weight matrix W1. h = x' * W = W[k,:] = v(input)
      RealVector h = W1.getRowVector(wordIndex); // 1xN row vector

      for (int contextWordIndex : wordPlusContexts.getSecond()) {
        Set<Integer> negativeContexts;
        if (sampleUnigram) {
          negativeContexts = negativeSampleContexts(wordIndex, noiseSampler);
        } else {
          negativeContexts = negativeSampleContexts(wordIndex);
        }
        // wordIndex is the input word
        // negativeContexts is the k negative contexts
        // contextWordIndex is 1 positive context

        // First update the output vectors for 1 positive context
        RealVector vPrime_j = W2.getColumnVector(contextWordIndex); // Nx1 column vector
        double u = h.dotProduct(vPrime_j); // u_j = vPrime(output) * v(input)
        double t_j = 1.0; // t_j := 1{j == contextWordIndex}
        double scale = sigmoid(u) - t_j;
        scale = eta * scale;
        RealVector gradientOut2Hidden = h.mapMultiply(scale);
        vPrime_j = vPrime_j.subtract(gradientOut2Hidden);
        W2.setColumnVector(contextWordIndex, vPrime_j);

        // Next backpropagate the error to the hidden layer and update the input vectors
        RealVector v_I = h;
        u = h.dotProduct(vPrime_j);
        scale = sigmoid(u) - t_j;
        scale = eta * scale;
        RealVector gradientHidden2In = vPrime_j.mapMultiply(scale);
        v_I = v_I.subtract(gradientHidden2In);
        h = v_I;
        W1.setRowVector(wordIndex, v_I);

        // Repeat update process for k negative contexts
        t_j = 0.0; // t_j := 1{j == contextWordIndex}
        for (int negContext : negativeContexts) {
          vPrime_j = W2.getColumnVector(negContext);
          u = h.dotProduct(vPrime_j);
          scale = sigmoid(u) - t_j;
          scale = eta * scale;
          gradientOut2Hidden = h.mapMultiply(scale);
          vPrime_j = vPrime_j.subtract(gradientOut2Hidden);
          W2.setColumnVector(negContext, vPrime_j);

          // Backpropagate the error to the hidden layer and update the input vectors
          v_I = h;
          u = h.dotProduct(vPrime_j);
          scale = sigmoid(u) - t_j;
          scale = eta * scale;
          gradientHidden2In = vPrime_j.mapMultiply(scale);
          v_I = v_I.subtract(gradientHidden2In);
          h = v_I;
          W1.setRowVector(wordIndex, v_I);
        }
      }
    }
  private static Intersection getIntersection(Ray ray, SphereObject obj, Model model) {
    RealMatrix transform = obj.getTransform();
    final RealMatrix transformInverse = obj.getTransformInverse();
    ray = ray.transform(transformInverse);
    Vector3D c = VectorUtils.toVector3D(obj.getCenter());
    Vector3D p0 = VectorUtils.toVector3D(ray.getP0());
    Vector3D p1 = VectorUtils.toVector3D(ray.getP1());
    float a = (float) p1.dotProduct(p1);
    Vector3D p0c = p0.subtract(c);
    float b = (float) p1.dotProduct(p0c) * 2.0f;

    float cc = (float) (p0c.dotProduct(p0c)) - obj.getSize() * obj.getSize();
    Double t = quadraticEquationRoot1(a, b, cc);
    if (t == null || t < EPSILON) {
      return new Intersection(false);
    }
    Intersection result = new Intersection(true);
    result.setObject(obj);
    final Vector3D p = p0.add(p1.scalarMultiply(t));
    RealVector pv = VectorUtils.toRealVector(p);
    pv.setEntry(3, 1.0);
    RealVector pt = transform.preMultiply(pv);
    result.setP(VectorUtils.toVector3D(pt));
    RealVector nv = pv.subtract(obj.getCenter());
    final RealVector nvt = transformInverse.transpose().preMultiply(nv);
    result.setN(VectorUtils.toVector3D(nvt).normalize());
    result.setDistance(t);
    return result;
  }