1
$\begingroup$

I was watching a video about policy gradients by Andrej Karpathy. At 10:00, it shows an equation for supervised learning for image classification.

$$\max\sum _{i} \log p(y_i \mid x_i)$$

I have worked with image classification models before, but I always minimized a cost function (aka loss function). I have also never seen someone maximizing a cost function for image classification in the wild.

  • So, what are the advantages of a minimizing loss function over a maximizing loss function in image classification?

  • Other than RL, which problems do we solve by maximizing a cost function?

$\endgroup$
1

1 Answer 1

1
$\begingroup$

There is really no difference between minimizing a cost function and maximizing a value function. One can be the reciprocal of the other, or the negative of the other, for example.

$\endgroup$
1
  • $\begingroup$ It is worth noting that the function that the OP gives is in fact the negative of the usual multi-class loss function. I would hazard a guess that Andrej used it to make a more direct comparison with RL. $\endgroup$ Oct 14, 2019 at 6:56

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .