?'S ON 5 STEPS OF SVM-KNN
----------------------------
1. How do we know that the the training examples involved in the SVM classification (i.e. the neighbors found by doing the preliminary K-NN) lie closely to the decision boundary?? (referencing section 2, the paragraph before they list the 5 steps of SVM-KNN)

2. Step 2 of SVM-KNN computes the "accurate" distance function on the K_sl SAMPLES... (are these called samples because we have pruned the neighbors using the crude distance in step 1, there is no random sampling going on, right?

3. I don't understand step 3, how can you compute the distance of a union of neighbors?

--------------------------------
?'S ON KERNELIZING THE DISTANCE
--------------------------------
4. I am unsure of how valid it is when they simply add the absolute value of eigenvalue of kernel matrix to the diagonal of the kernel matrix to make it positive-definite.. it seems so sketchy, yet I am still getting familiar with the kernel matrix, and all the operations with it =).. plus, I don't know what it means to strengthen self-similarity? How can something become more similar to itself?

----------------------------
QUESTION ON RESULTS COMMENT
----------------------------
5. Paragraph 3 of 4.1.. What does it mean that a number of workarounds were needed to compute the shape context based distance?

Thank you! =)
-----------------

Just a comment.

I really liked how they based their approach on an interpretation of biological vision where people classify objects based on similarity to prototypes in miliseconds and then slowly determine what the specific object is. One thing that stood out was how much results depend on the shape and texture distances used, finding the appropriate distance seems almost as important as finding the right classifier.

Can you explain a bit about how DAGSVM works (especially as it relates to the multi-class problem)?

About non-PSD SVM kernel, is it really a good solution to add the diagnal regularization term? Because there're no good accurate explaination in feature space.

In my past experience, sometimes the weight of regularization term also affects the classificatoin accuracy.

This paper is an example of a hybrid of two algorithms performing better than each one alone on certain datasets.

Is there something particularly attractive about the pairing of SVMs with KNN, or could this paper have just as easily been about "<insert some other classifier here>-KNN"?