Hello ,
if i have a trained neural network with softmax function for classification.
During the testing phase (prediction), can i use argmax function instead of softmax function ??!
For example, if the output of last layer before the softmax function is [2,4,2,1]. In case argmax function, the output will be [0,1,0,0] and i am looking for the largest value in my application.
This will help to reduce the complexity during the implementation process.
The usage of Softmax is needed in a neural network structure (as opposed to argmax) because it is continuous, and therefore differentiable. The argmax function is not. (https://math.stackexchange.com/questions/2354134/is-text-argmax-fx-y-a-differentiable-function-of-x)
If you are interested in a sparse vector for output you could add another layer after softmax which would hard set the largest value from the softmax layer to 1 and all other values to zero. This would need to be done after training.
Thank you for your response.
I know the softmax is needed during the training. But i asked about testing phase (after training ), could i remove softmax layer and use argmax ??? Alexander Atkinson
I mentioned that in the second half of my answer. Yes Abdelhay Ali, you can add another layer after training which would convert the output of the softmax layer into an argmax for sparse vector output.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training