User Antoni Parellada had a long derivation here on logistic loss gradient in scalar form. Using the matrix notation, the derivation will be much concise. Can I have a matrix form derivation on logistic loss? Where how to show the gradient of the logistic loss is

$$ A^\top\left( \text{sigmoid}~(Ax)-b\right) $$

  1. For comparison, for linear regression $\text{minimize}~\|Ax-b\|^2$, the gradient is $2A^\top\left(Ax-b\right)$, I have a derivation here.

  2. Related question: Matrix notation for logistic regression

  • $\begingroup$ As credited in the linked answer, the credit for the derivation goes to the students' cumulative notes within the excellent course by Andrew Ng on ML. $\endgroup$ May 11 '17 at 2:24
  • 1
    $\begingroup$ Why not try translating the scalar-form derivation as a self-study? (BTW you should edit the question to explicitly state the loss itself, in your preferred notation, to make it self contained) $\endgroup$
    – GeoMatt22
    May 11 '17 at 2:36

Here is my try

$$J(x) = -\frac{1}{m}\sum_{i = 1}^{m} b_iln(h_i) + (1 - b_i)ln(1 - h_i)$$

where $h_i = \sigma(x^Ta_i)$. Let $A = [a_1^T, \dots, a_m^T]^T$. Assuming $ln, \sigma, \frac{1}{\cdot}$ work element-wise on vectors, $\odot$ is element-wise multiplication and $\mathbb{1}$ is a vector of $1$s we have

$$J(x) = -\frac{1}{m}\big[b^Tln(\sigma(Ax)) + (\mathbb{1} - b)^Tln(\mathbb{1} - \sigma(Ax))\big]$$


$$\frac{\partial J(x)}{\partial x} = -\frac{1}{m}\Big[\frac{\partial}{\partial x}b^Tln(\sigma(Ax)) + \frac{\partial}{\partial x}(\mathbb{1} - b)^Tln(\mathbb{1} - \sigma(Ax))\Big] \\ = -\frac{1}{m}\Big[\frac{\partial ln(\sigma(Ax))}{\partial x}b + \frac{\partial ln(\mathbb{1} - \sigma(Ax))}{\partial x}(\mathbb{1} - b)\Big] \\ = -\frac{1}{m}\Big[\frac{\partial \sigma(Ax)}{\partial x} \big(\frac{1}{\sigma(Ax)}\odot b\big) + \frac{\partial \mathbb{1} - \sigma(Ax)}{\partial x}\big(\frac{1}{\mathbb{1} - \sigma(Ax)}\odot (\mathbb{1} - b)\big)\Big] \\ = -\frac{1}{m}\Big[\frac{\partial Ax}{\partial x} \big(\sigma(Ax) \odot (\mathbb{1} - \sigma(Ax)) \odot \frac{1}{\sigma(Ax)}\odot b\big) - \frac{\partial Ax}{\partial x}\big(\sigma(Ax) \odot (\mathbb{1} - \sigma(Ax)) \odot \frac{1}{\mathbb{1} - \sigma(Ax)}\odot (\mathbb{1} - b)\big)\Big] \\ = -\frac{1}{m}\Big[A^T \big((\mathbb{1} - \sigma(Ax)) \odot b\big) - A^T\big(\sigma(Ax)) \odot (\mathbb{1} - b)\big)\Big] \\ = -\frac{1}{m}\Big[A^T \big(b - \sigma(Ax) \odot b - \sigma(Ax) + \sigma(Ax) \odot b\big)\Big] \\ = -\frac{1}{m}\Big[A^T \big(b - \sigma(Ax)\big)\Big] \\ = \frac{1}{m}\Big[A^T \big(\sigma(Ax) - b\big)\Big] $$


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.