Clarity for quiz question

Course Q&A Deep Learning Specialization Structuring Machine Learning Projects
1

There was question in the quiz of the week 2:

Your goal is to detect road signs (stop sign, pedestrian crossing sign, construction ahead sign) and traffic signals (red and green lights) in images. The goal is to recognize which of these objects appear in each image. You plan to use a deep neural network with ReLU units in the hidden layers. For the output layer, which of the following gives you the most appropriate activation function?

Out of the given options, I chose “SoftMax”, but that was incorrect. I am not able to understand the reason.

2

Right, softmax would output only one possibility, you need many outputs that are present not only one!

3

So which activation function is the right choice: Sigmoid, ReLu or Linear Regression?

4

You tell me, since you need to predict the presence or absence of multiple objects in each image, which activation function allows each output neuron to independently represent the probability of the corresponding object being present in the image?

1 Like
5

To me, Sigmoid can be used. The Sigmoid activation function squashes the output of each neuron to the range [0, 1], representing the probability of the corresponding object being present. Each neuron in the output layer can independently represent the probability of its corresponding object being present or absent in the image.

1 Like
6

But I am not sure

7

That sounds plausible to me. There’s one way to find out if the quiz grader agrees or not. :nerd_face:

8

Softmax only re-scales multiple outputs so their sum is 1. But since you want to identify multiple items in each image, the sum of these independent classes won’t necessarily be 1.

Also:

  • ReLU isn’t used as an output activation.
  • Linear Regression isn’t used for classification.
1 Like