In gradient descent, I know that local minima occur when the derivative of a function is zero, but when the loss function is used, the derivative is equal to zero only when the output and the predicted output are the same (according to the equation below).

So, when the predicted output equals the output, that means the global minima is reached! So, my question is: How can a local minima occur, if zero gradient occurs only for the "perfect" fit?

$$\theta_j := \theta_j - {\alpha \over m} \sum_{i=1}^M (\hat y^i-y^i)x_j^i$$


The equation you used for gradient descent isn't general; it's specific for linear regression.
In linear regression, there is indeed only a single global minimum and no local minima; but for more complex models, the loss function is more complex, and local minima are possible.

  • $\begingroup$ but in backpropagation(neural network), also derivative equals to zero when expected output=target as in equation(non-linear sigmoid function): visualstudiomagazine.com/articles/2017/06/01/~/media/ECG/… Does this case also have single global minima? $\endgroup$ – AI_new2 Jul 27 '20 at 10:07
  • $\begingroup$ output=target for each term in the sum is not the only way that the sum can equal zero. $\endgroup$ – Itamar Mushkin Jul 27 '20 at 10:14
  • $\begingroup$ Suppose it a single example for training will lead to having single minima, so in stochastic gradient descent, there will be no local minima! $\endgroup$ – AI_new2 Jul 27 '20 at 10:42
  • $\begingroup$ A single example can't be used for training if there's more than one parameter in the model... $\endgroup$ – Itamar Mushkin Jul 29 '20 at 5:56

The premise of “no minimum without a perfect fit” is incorrect.

Let's look at a simple example with square loss.

$$L(\hat{y}, y) = \sum_i (y_i-\hat{y}_i)^2$$

$$ (x_1, y_1) = (0,1)$$ $$ (x_2, y_2) = (1,2)$$ $$ (x_3, y_3) = (3,3)$$

We decide to model this with a line: $\hat{y}_i = \beta_0 + \beta_1 x_i$.

Let's optimize the parameters according to the loss function.

$$L(\hat{y}, y) = (1-(\beta_0 + \beta_1(0)))^2 + (2-(\beta_0 + \beta_1(1)))^2 + (3-(\beta_0 + \beta_1(3)))^2$$

Now we take the partial derivatives of $L$ with respect to $\beta_0$ and $\beta_1$ and do the usual calculus of minimization.

So we minimize the loss function, but we certainly do not have a perfect fit with a line.

  • $\begingroup$ This loss function does not have any local minima though... $\endgroup$ – Itamar Mushkin Jul 28 '20 at 18:14
  • $\begingroup$ The critical point is $(\hat{\beta}_0, \hat{beta}_1) = (\frac{8}{7}, \frac{9}{14})$. The eigenvalues of the Hessian matrix at $(\frac{8}{7}, \frac{9}{14})$ are $13\pm\sqrt{113}$, both of which are $>0$, making the critical point a minimum. $\endgroup$ – Dave Jul 28 '20 at 22:31
  • $\begingroup$ Yes, of course it has a minimum - but it's a global minimum. OP asked about local minima other than the global minimum. $\endgroup$ – Itamar Mushkin Jul 29 '20 at 5:55
  • 1
    $\begingroup$ The idea that the gradient is zero only for the perfect fit is incorrect. $\endgroup$ – Dave Jul 29 '20 at 10:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.