I have a Keras NN model where I apply the Monte Carlo dropout approach as a predictive method to evaluate the uncertainty of the model outputs. From my research in the probabilistic neural networks, I understand that the lower the entropy (close to 0), the more probable that the predicted value is really certain.

In my model, I ran the inference step 100 times, averaged model output probabilities, and calculated the Entropy. The following is a randomly chosen instance from my results (mean_proba is averaged over 100 iterations of predict):

instance_id       y_true   mean_proba     Entropy
    2301            1        0.9009       0.0940

My question is: from the instance# 2301 result, does this imply that the model has a low uncertainty, i.e., high information gain, that instance# 2301 should really be predicted as 1? Is this interpretation sound?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.