How does one unit/neuron know what it needs to calculate?

Course Q&A Machine Learning Specialization Advanced Learning Algorithms
1

Capture
Capture1019Ă—508 227 KB

How each neuron identifies what it needs to calculate. For example, in this layer how first neuron knows it need to calculate affordability?
And as we use value of weights where cost function is minimum then why each neuron in a layer has different weights?

1 Like
2

Hi @Hammad_Hassan61 ,

The architect of the network is responsible for making decision according to the business’s requirement. In this case, the model is required to make prediction on the demand of a product. So the designer will need to consider what are the factors that would influence the demands, and design the neuron accordingly.

2 Likes
3

The neuron doesn’t actually identify what quality (ex. affordability) that it is calculating. All it does is minimize the costs by adjusting the weights during training.

With that said, once a neural network has been trained, we (as human researchers) can look at the neurons and try to figure out that association. That’s what this diagram is showing. Note that this is a simple example (to help with your intuition), and in practice, we may or may not be able to actually make this association.

The weights are different because each neuron should contribute to the cost differently. The programming exercises should help with your understanding of this (in case you haven’t done them yet).

2 Likes
4

How can the neuron be designed? In the next lecture focusing on the same problem, the instructor will utilize the sigmoid function for all neurons in the layer to calculate the activation values. If the same function is used with different weights, how can the output factors (affordability, quality etc) be different for each neuron?

1 Like
5

Hi @Hammad_Hassan61 ,

Each neuron is designed to extract certain information. So, you may have a neuron focused on affordability, for example. The weights for each neuron in this case is randomly generated, so the output of each neuron is not identical. Applying the activation function is to range the value between 0 and 1.

2 Likes
6

It doesn’t know this. The features are just numbers. The neuron units have no idea what they represent.

Each unit only knows that its job is to use the weight that minimizes the cost