Hello
For our project, our task is to decode fear and quantify it from neural recordings in rodents. We will do this by feeding colormaps (from the data) to a CNN. I have a question about the label generation in CNNs. Is there any difference in model performance if the labels we select are class numbers from the very start, or if we only classify in the last layer of the NN based on the numerical prediction after doing a regression step? Formulated differently; do we make our labels discrete class numbers directly, or continuous numbers and classify only in the last layer? Is there a standard procedure for this problem or does it depend on the application?
Thanks in advance!
I will try to answer even though the question is a little bit unclear to me. In general, you use a CNN to extract useful representations then feed these representations to a classifier/regressor. For classification, the labels are encoded as probabilities by one-hot encoding them and interpreting the output as the probability that your input is in a given class.
I hope I answered your question, otherwise try to reformulate it.
generate labels for CNN
Hello
For our project, our task is to decode fear and quantify it from neural recordings in rodents. We will do this by feeding colormaps (from the data) to a CNN. I have a question about the label generation in CNNs. Is there any difference in model performance if the labels we select are class numbers from the very start, or if we only classify in the last layer of the NN based on the numerical prediction after doing a regression step? Formulated differently; do we make our labels discrete class numbers directly, or continuous numbers and classify only in the last layer? Is there a standard procedure for this problem or does it depend on the application?
Thanks in advance!
I will try to answer even though the question is a little bit unclear to me. In general, you use a CNN to extract useful representations then feed these representations to a classifier/regressor. For classification, the labels are encoded as probabilities by one-hot encoding them and interpreting the output as the probability that your input is in a given class.
I hope I answered your question, otherwise try to reformulate it.
Add comment