This is perfect! I'm so sick of all these fancy literatury stuff from professors all over the world who can only communicate through differential equations. THIS is how it should be explained. Thank you good sir!
Pretty good explanation but you never showed what happens if the number of K you are searching for is bigger than the number of points in the specific area. For example let's say you have a new point in R4 which has 3 points and you are searching for 4-NN for that point. Thank you again for this video, really liked it
Doesn't answer your question directly, but in FAISS IVF index, if k is more than number of items in a cell, it returns -1 id for the extra required neighbors, solution is to increase default nprobe=1 to probe more cells.
3:56 I thought that a kdtree can search nearest neighbor in logn and delete or add a point in logn so k nearest neighbors could be considered klogn which is less than n
Thanks for a great video! One questions, @9:23 new point we check if given point is below or above the blue line. The way you recognize whether point is above or below is by calculating distance between (point, 1) and (point, 9) ?
What's your qualification? Somehow I cannot find any information about your education etc. Awesome videos by the way, a lot easier to understand than what every professor tries to explain.
Great video as always Ritvik. Am I correct that building the tree is an O(N) operation? That is, if I have only one new data point and haven't yet constructed the tree, will this still save any time over the exhaustive method? If not, then I presume building a forest would imply some break even point. Thanks.
Here is what I think: each region has two points. So use a metrics (e.g. distance) from this given new point to the begin and to the end point and go with the closer one. The closeness can be Euclidean distance, or Cosine distance, or some other metrices.
Such a nice recursive challenge. anyone have an idea how to define a function to recursivley solve this kind of algorithm, given a creiteria of maximum points?
![[CVPR20 Tutorial] Billion-scale Approximate Nearest Neighbor Search](/img/n.gif)
Wow you have no idea how much i needed this for my current work project. Thanks as always for a fantastic explanation
I have implemented ANN on my own after watching your video. Thanks for the great explanation ritvik
This is perfect!
I'm so sick of all these fancy literatury stuff from professors all over the world who can only communicate through differential equations. THIS is how it should be explained. Thank you good sir!
Thank you so much sir this explanation shows your exceptional ability to teach. So enlightening!
Both formats are cool
OMG. I hope all my lecturers will explain that clearly and intuitively. Thankss
Clear explanation and very resourceful!
This is brilliant! Thank you so much for showing us this method!
The lesson was clear and paper can be easier for you to control and work with. So this is fine. Thank you for the lesson!
You have a very clear but not too wordy style. *SUBSCRIBED*
Pretty good explanation but you never showed what happens if the number of K you are searching for is bigger than the number of points in the specific area.
For example let's say you have a new point in R4 which has 3 points and you are searching for 4-NN for that point.
Thank you again for this video, really liked it
Doesn't answer your question directly, but in FAISS IVF index, if k is more than number of items in a cell, it returns -1 id for the extra required neighbors, solution is to increase default nprobe=1 to probe more cells.
Mate you really know how to explain things. Thanks for your time and dedication.
This format is better. Thanx.
Excellent Video
Glad you enjoyed it!
Thank you😊
Thank you so much for the simple and clear explanation with examples!
You're very welcome!
Great explanation!
Glad it was helpful!
Thanks for sharing such a detaild and thorough explanation!
My pleasure!
Very clear explanation! I think I got it in one pass! Pace is good. Thanks! (PS. the paper format is fine!)
THIS WAS AMAZING!!!!!!!!!!!!!!!
I really like this format for this kind of explanation
Like explainnig how a technique works
very good vid, thanks
Glad you liked it!
best explanation ever. thank you
thank you very much 🙏🏼
Of course!
Thank you so much for a beautiful lesson. Reminded me of my elementary school days and how teachers used to teach back then.
Thanks! Good vid :)
Greatly explained
Perfect explanation! Thanks :D
thank you very much, it was so helpful
Very well explained!!
Glad you think so!
3:56 I thought that a kdtree can search nearest neighbor in logn and delete or add a point in logn so k nearest neighbors could be considered klogn which is less than n
well explained! thanks!
"Lowest Complexity for Knn is O(n)" is not True!!
Using kd-tree the complexity becomes
O(log n).
I was also thinking about kd tree and ball tree used in sklearn... Are you aware of any other methods??
@@jasdeepsinghgrover2470 LSH
Thanks for sharing... Will learn more about it as well
Thanks for a great video! One questions, @9:23 new point we check if given point is below or above the blue line. The way you recognize whether point is above or below is by calculating distance between (point, 1) and (point, 9) ?
I now wonder if this is a sensible algorithm for collision detection
Really cool :O thank you
How would we determine that a point is above and below a line using code ?
What's your qualification? Somehow I cannot find any information about your education etc. Awesome videos by the way, a lot easier to understand than what every professor tries to explain.
such a great explanation! Wonder do you also have a similar video for HNSW? Thanks!
Great video as always Ritvik.
Am I correct that building the tree is an O(N) operation? That is, if I have only one new data point and haven't yet constructed the tree, will this still save any time over the exhaustive method?
If not, then I presume building a forest would imply some break even point.
Thanks.
Thanks for this excellent video! Is there a poplar library that helps to experiment with ANN on local machine for a small set of data?
I like it MUCH better. I found it sometimes overwhelming to be confronted with all the info and not yet have an explanation.
Paper is better, I think. Moving the papers around is like zooming without moving the camera.
I still don't understand how do you classify the new point? region wise or is there any other method?
Here is what I think: each region has two points. So use a metrics (e.g. distance) from this given new point to the begin and to the end point and go with the closer one. The closeness can be Euclidean distance, or Cosine distance, or some other metrices.
Such a nice recursive challenge. anyone have an idea how to define a function to recursivley solve this kind of algorithm, given a creiteria of maximum points?
Looks like a sort of a binary search
Is ANNOY using Voronoi ?