Active Learning | Tutorial on Active Learning from Theory to Practice | ICML

[@beauzeta1342]

Scientific talk starts at 11:14

[@jeroenritmeester73]

Passive learning is like watching KZbin's recommended videos, sometimes learning something new. Active learning is watching lectures on new technologies, ensuring you encounter something you haven't seen before.

[@frederikzilly7998]

Thank you for the illustrative talk. Just one question: In minute 22:00 there is the example of a probability p of y, given x p(y|x) potentially being below 0. Should it be better replaced by a threshold t ? I just stumbled seeing that negative probability...

[@DrOsbert]

Hello from the Editor of AIP, From what I understand, the talk is mostly referred to a generic binary classification problem, so for the case that probability, p(y|x) below 0, instances are classified as negative / -1 / 0 , and vice versa. Here, the threshold, t is 0, which is the default threshold and it may not represent an optimal interpretation of the predicted probabilities. (From Machine Learning Mastery machinelearningmastery.com/threshold-moving-for-imbalanced-classification/) This might be the case for a number of reasons, such as: 1. The predicted probabilities are not calibrated, e.g. those predicted by an SVM or decision tree. 2. The metric used to train the model is different from the metric used to evaluate a final model. 3. The class distribution is severely skewed. 4. The cost of one type of misclassification is more important than another type of misclassification. Therefore, the 0 threshold value can be replaced by any value depending on your dataset, domain and use cases. I think, for illustration purpose and digest-able, the presenter make a hard threshold of > 0 for positive instances. However, I might be wrong, if you have further doubt, I can help you to reach out the author for further clarification? Disclaimer: The above comments are solely my personal opinion.

[@manohariisc]

@@DrOsbert p(y|x) becoming negative is problematic....are we not violating one of the axioms?

[@zoeyx1998]

Thanks! This helps a lot.

[@surflaweb]

only some cats can understand this

[@BigAsciiHappyStar]

8 Out Of 10 Cats Does Active Learning (please don't delete me) 😊

[@jakebrowning2373]

OMG