From e3110daaf7027a0f5986b332ed140c11624c446b Mon Sep 17 00:00:00 2001 From: Eric Zelikman Date: Sat, 10 Oct 2020 00:13:32 -0400 Subject: [PATCH] likelihood -> rephrased --- C1W1_(Colab)_Inputs_to_a_pre_trained_GAN.ipynb | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/C1W1_(Colab)_Inputs_to_a_pre_trained_GAN.ipynb b/C1W1_(Colab)_Inputs_to_a_pre_trained_GAN.ipynb index 42b765b..eb9c103 100644 --- a/C1W1_(Colab)_Inputs_to_a_pre_trained_GAN.ipynb +++ b/C1W1_(Colab)_Inputs_to_a_pre_trained_GAN.ipynb @@ -78,7 +78,7 @@ "\n", "> *Image Credit: [Google](https://about.google/stories/machine-learning-qa/)*\n", "\n", - "In terms of probabilities, the classifier wants to find $p(y|x)$, or the likelihood that given an image input $x$, you want to know the probability of its class $y$ being cat, $p(y=\\text{cat}|x)$, or dog, $p(y=\\text{dog}|x)$.\n", + "In terms of probabilities, the classifier wants to find $p(y|x)$: the probability that given an image input $x$, the image class $y$ is cat, $p(y=\\text{cat}|x)$, or dog, $p(y=\\text{dog}|x)$.\n", "\n", "The discriminator is simply a classifier with two classes: real and fake. Given an input x, the discriminator will calculate the probabilities $p(y=\\text{real}|x)$ and $p(y=\\text{fake}|x)$ and classify $x$. The input $x$ can be anything that you have the generator create and is not limited to images. Your GAN can be trained on videos, text, audio, etc.\n" ] @@ -93,7 +93,7 @@ "\n", "Generators are designed to have a different goal from discriminators (classifiers). Imagine you're working at a tropical fruit stand and asked to sort the fruit into two categories: coconuts and starfruit. That's the job of a classifier. But what if a customer comes up to the stand, and asks: what is a starfruit? You can't just say that it's not a coconut. You would need to explain what makes something a starfruit and what doesn't, not just its differences from a coconut. That's the job of a generator: to represent different classes in general, not just distinguish them. \n", "\n", - "In terms of probabilities, the generator wants to figure out $p(x|y)$, or the likelihood that given the fact that you want to generate a starfruit $(y=\\text{starfruit})$, what the resulting image $(x)$ could possibly look like. The output space of possible starfruit images is huge, so that makes this challenging for the generator.\n", + "In terms of probabilities, the generator wants to figure out $p(x|y)$: the probability that, given that you generated a starfruit $(y=\\text{starfruit})$, the resulting image $(x)$ is the one generated. The output space of possible starfruit images is huge, so that makes this challenging for the generator.\n", "\n", "This can be a much harder task than discrimination. Typically, you will need the generator to take multiple steps to improve itself for every step the discriminator takes. It's easy to tell the difference between a coconut and a starfruit when you look at a mix of them, but to know exactly all the features of all possible coconuts in the world? That's a lot, but it's really cool if you can get even close to it, because you can start generating all sorts of coconuts and starfruit when you do. I don't have a generator for these tropical fruits figured out for you to play with, but you can generate all kinds of cool things with these models, like realistic faces!\n", "\n",