0
$\begingroup$

Let's say I have a neural net with Dense layers. The input layer has 3 neurons, the single hidden layer has 5 neurons, and the final output layer has 2 neurons.

For layer 1, 3 inputs go in and 5 inputs go out. For layer 2, the 5 inputs from layer 1 go in and 2 inputs go out. So for layer 3, would 2 inputs go in and 2 inputs come out?

$\endgroup$
2
$\begingroup$

In a Dense NN,

  • Input to a Layer depends on the output of the previous layer and its Neuron count
  • Previous Layer for the first layer is the Input Data Features
  • Output of a Layer is equal to the Layer's neuron count. A copy of each goes to all the Neuron of the Next Layer

It means for your case,
- First Layer will have *M(Input features)3 inputs going in i.e. M to each Neuron and 3 coming out.
- The Second layer will have *3(Previous Layer)5 inputs going in and 5 coming out
- The Last layer will have 5(Previous Layer)*2 inputs going in and 2 coming out


A Neural Network is a complex Tensor operation. Arrow and Circle are logical representations. Each arrow got a weight and each Circle got the Activation function and the Bias term. So, you may count this way too.

$\hspace{3cm}$enter image description here
Validate this with Keras Model.summary result which shows the parameter count.

$\endgroup$
0
$\begingroup$

No - the input and outputs sizes are independent from one another in a deep full-connected network. You could e.g. have input matrix shape (100, 100, 100) and output shape (1,).

$\endgroup$
0
$\begingroup$

I think you are confusing weights and neurons a bit.

Output of the last layer: problem-specific (binary: single neuron, multiclass/multilabel: number of neurons=number of Classes) number of neurons.

Input of the last layer: weights equal to the number of neurons in the previous layer + 1 (bias) per each output neuron.

$\endgroup$

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.