Consider, there are two classes of data and we have learned the SVM parameters in terms Lagrange multipliers. There are many learning techniques to learn these parameters quadratic programming or sequential minimal optimization etc.
The question is, once we have learned our classifier, can we update the parameters somehow, once we get a new unknown sample . We can easily pass the unknown sample through the classifier and get a result.
- But is there a way to learn something blindly from the sample?
- This is essential for our case: in our application the input to the classifier changes slowly over time. So we need to capture the slow changes somehow blindly. Is it possible?
The list of answers from Is it possible to append training data to existing SVM models? are related to how to train when you get a new sample. Kindly note in that case you know the label of the new sample, in other words the class of the new sample is known or using oracle. The question here refers to the case when the class of the sample is not known As such in my opinion this is not duplicate. This question may not have an answer though. The closet approach so far seen is https://en.wikipedia.org/wiki/Active_learning_(machine_learning) active learning, however this not answer the original question of slow changing. To be more specific some active learning considers support vectors and weights as a metric to decide a label but here question is related to change of support vectors to cope with slow changes.
Let me write other way, I am looking for active learning without knowing the label. All those references deal with how to choose a sample, the oracle or somebody would specify the class of the sample.