0:00:16 | i i think you uh many was out slow be presenting |
---|
0:00:19 | uh paper |
---|
0:00:21 | and title the the group |
---|
0:00:23 | two two of for or as i don't because this problem |
---|
0:00:25 | uh by brian kind of a my adviser |
---|
0:00:27 | uh fess with peter image |
---|
0:00:32 | okay so um a well known problem that's been around a while is uh your a title for "'cause" this |
---|
0:00:37 | problem |
---|
0:00:37 | where we uh C to identify |
---|
0:00:40 | and orthogonal cycle transformation between two data matrices |
---|
0:00:43 | and uh |
---|
0:00:44 | this context we seek to |
---|
0:00:46 | minimize the frobenius norm squared |
---|
0:00:49 | of the difference of a |
---|
0:00:51 | in this case |
---|
0:00:52 | the rows of B with rotated rows of a |
---|
0:00:56 | and and this is useful uh |
---|
0:00:58 | uh it to serve as as a as a geometric |
---|
0:01:02 | uh transformation or a geometric rotation to to correct for |
---|
0:01:06 | uh |
---|
0:01:07 | rotational distortion as |
---|
0:01:08 | uh distortions between |
---|
0:01:10 | i mitch each a and B |
---|
0:01:13 | um |
---|
0:01:14 | uh |
---|
0:01:15 | going going to |
---|
0:01:16 | going to a second order version uh there was there's the two sided or because he's problem |
---|
0:01:21 | or you have to uh |
---|
0:01:23 | symmetric |
---|
0:01:25 | similarity matrices C and C are little are can to give is |
---|
0:01:28 | reference |
---|
0:01:29 | and |
---|
0:01:31 | in this in this problem we have |
---|
0:01:33 | rather than um |
---|
0:01:36 | uh |
---|
0:01:36 | de de de rotation during on one side we have it of on both sides |
---|
0:01:40 | and in the form are transpose |
---|
0:01:42 | C R and and again we're trying to match up |
---|
0:01:45 | um this this this to side rotation with with |
---|
0:01:48 | with some |
---|
0:01:50 | uh |
---|
0:01:51 | some reference |
---|
0:01:52 | uh |
---|
0:01:53 | similarity so |
---|
0:01:54 | uh |
---|
0:01:56 | the says a bunch of applications uh |
---|
0:01:59 | one being a a covariance or or correlation |
---|
0:02:02 | a matrix matching so |
---|
0:02:04 | we eight |
---|
0:02:05 | a data matrix to we've induced |
---|
0:02:08 | uh a double rotation |
---|
0:02:10 | on the covariance matrix |
---|
0:02:12 | um a more interesting problem |
---|
0:02:14 | would be a related to a graph matching problem so in a ice metric graph matching |
---|
0:02:20 | we're given uh two adjacency matrices |
---|
0:02:22 | again C and C R |
---|
0:02:24 | and uh |
---|
0:02:25 | this can be weighted |
---|
0:02:26 | um |
---|
0:02:27 | and and and and the at a is to find a a permutation matrix |
---|
0:02:32 | which |
---|
0:02:33 | best uh |
---|
0:02:35 | which we we which serves to to reorder the the nodes of one graph and to the other man in |
---|
0:02:41 | in the other graph |
---|
0:02:42 | uh again |
---|
0:02:43 | matching in in in the for been used um score different so |
---|
0:02:47 | this is a hard problem big |
---|
0:02:48 | given this combinatorial nature |
---|
0:02:51 | uh there are |
---|
0:02:53 | one one one popular method is |
---|
0:02:56 | when me i'm as method um and |
---|
0:02:58 | involves relaxing this problem |
---|
0:03:00 | to |
---|
0:03:01 | an orthogonal |
---|
0:03:03 | uh constraint |
---|
0:03:05 | and then and then finding the closest permutation matrix |
---|
0:03:08 | as the hungarian algorithm |
---|
0:03:13 | okay so if we go back to the of that don't is problem the set this has a known solution |
---|
0:03:17 | and um |
---|
0:03:19 | given |
---|
0:03:20 | uh the the the right circumstances |
---|
0:03:22 | it's a unique and this given by D |
---|
0:03:24 | uh |
---|
0:03:25 | single value decomposition of A transpose be where |
---|
0:03:29 | a singular values |
---|
0:03:30 | um |
---|
0:03:30 | have to be |
---|
0:03:31 | transform just a |
---|
0:03:33 | all ones |
---|
0:03:34 | oh have for the two sided however |
---|
0:03:36 | it's a it's |
---|
0:03:38 | it's |
---|
0:03:38 | a bit more complicated because |
---|
0:03:40 | uh we have a close form solution but it's really family of solutions so |
---|
0:03:44 | if the eigen decomposition of C is is V |
---|
0:03:47 | lamb V transpose and C are B R and R |
---|
0:03:51 | the are transpose |
---|
0:03:52 | then |
---|
0:03:53 | uh |
---|
0:03:54 | a solution to this problem is given by |
---|
0:03:57 | V V D V are transpose where D is a a is a diagonal matrix where each i an entry |
---|
0:04:02 | is plus one or minus one |
---|
0:04:04 | so |
---|
0:04:06 | um |
---|
0:04:07 | right away we see that we can have know |
---|
0:04:09 | to the N possible solutions |
---|
0:04:11 | at least |
---|
0:04:14 | uh so so so i was a bit of a history um |
---|
0:04:17 | in in in this talk we're going to consider |
---|
0:04:21 | scenario area where |
---|
0:04:22 | a for example a graph verdi sees don't just have |
---|
0:04:26 | similarity measures between nodes but |
---|
0:04:29 | um but also |
---|
0:04:31 | there some side information |
---|
0:04:33 | and and so this paper considers |
---|
0:04:35 | case where |
---|
0:04:36 | the a the seize |
---|
0:04:37 | our are parsley later |
---|
0:04:38 | in to go |
---|
0:04:40 | uh so |
---|
0:04:41 | so for example if uh |
---|
0:04:43 | let's say were modeling a face so |
---|
0:04:46 | so you know certain certain notes belong to the i certain nodes belong to the nose set |
---|
0:04:50 | um or we can think of a a spatial configuration so |
---|
0:04:54 | and uh in uh |
---|
0:04:56 | fmri analysis we might have |
---|
0:04:58 | certain regions of interest |
---|
0:04:59 | so |
---|
0:05:00 | we have kind activity is between |
---|
0:05:02 | of voxels but we also know uh what are our why those walks |
---|
0:05:05 | two |
---|
0:05:09 | okay |
---|
0:05:10 | so suppose |
---|
0:05:11 | are are are data dataset |
---|
0:05:12 | is parsley did in |
---|
0:05:14 | capital "'em" groups |
---|
0:05:16 | but is for each point i there's |
---|
0:05:18 | there's some group that i belongs to and that's uh |
---|
0:05:21 | will note that G of i |
---|
0:05:24 | and uh uh uh the goal is |
---|
0:05:26 | to |
---|
0:05:27 | again |
---|
0:05:28 | estimate an R |
---|
0:05:30 | uh a in in in in the two sided |
---|
0:05:32 | of course this problem but |
---|
0:05:33 | here are is gonna be block diagonal |
---|
0:05:36 | and and and those block diagonals correspond to the |
---|
0:05:40 | nature of of the uh |
---|
0:05:41 | problem |
---|
0:05:43 | so uh because problem the uh |
---|
0:05:46 | groups to sided or that process problem and uh |
---|
0:05:50 | of course when when when M equals one we |
---|
0:05:52 | go back to the original |
---|
0:05:54 | a standard to prior |
---|
0:05:59 | so |
---|
0:06:00 | get |
---|
0:06:01 | given T block tan'll constraint |
---|
0:06:02 | um |
---|
0:06:04 | it essentially introduces |
---|
0:06:06 | some some linear constraints that is the elements of |
---|
0:06:10 | this large matrix R |
---|
0:06:12 | a a such that |
---|
0:06:14 | uh |
---|
0:06:15 | the uh it in X P Q zero if um |
---|
0:06:19 | if uh |
---|
0:06:21 | notes P and Q of the |
---|
0:06:25 | and uh some some benefits of this approach are |
---|
0:06:28 | you know the the |
---|
0:06:29 | these constraints regular as the problem so |
---|
0:06:32 | uh we might be reducing the effects of fitting |
---|
0:06:36 | and also were were we're using prior knowledge that uh |
---|
0:06:39 | we know about a about this graph |
---|
0:06:42 | um |
---|
0:06:43 | also we have uh |
---|
0:06:45 | a few were |
---|
0:06:46 | parameters to estimate so it's uh |
---|
0:06:49 | it's a |
---|
0:06:50 | were dealing with a i |
---|
0:06:52 | a a a a a a lower dimensional problem |
---|
0:06:56 | okay so so for "'em" greater than one we so that it it it does have a |
---|
0:07:01 | a close form family |
---|
0:07:02 | solutions |
---|
0:07:03 | but for "'em" greater than one |
---|
0:07:05 | where we're gonna have to look at an iterative |
---|
0:07:07 | uh technique |
---|
0:07:08 | so that let see i i and C I R denote be with an region or within groups similarity structures |
---|
0:07:14 | uh and |
---|
0:07:16 | see i J where i is not equal to J we'll what we'll do not the cross |
---|
0:07:21 | somewhere structures |
---|
0:07:22 | so |
---|
0:07:23 | again given that this is uh going to be an iterative |
---|
0:07:26 | solution |
---|
0:07:27 | uh we'd like to start from |
---|
0:07:29 | a a a a a good starting point |
---|
0:07:31 | and so |
---|
0:07:32 | a logical starting point would be to just consider the you within group similar |
---|
0:07:38 | so we can start it |
---|
0:07:39 | each block |
---|
0:07:40 | element of our |
---|
0:07:42 | rotation matrix |
---|
0:07:43 | to be |
---|
0:07:44 | uh V M M D M V M M R transpose |
---|
0:07:48 | where |
---|
0:07:49 | uh |
---|
0:07:50 | the V and the Rs |
---|
0:07:52 | correspond to the eigen decompositions of |
---|
0:07:55 | oh of uh |
---|
0:07:56 | of uh a C M M and |
---|
0:07:58 | and and M R |
---|
0:08:02 | so and and so from the starting point our goal is to improve on this initial estimate |
---|
0:08:06 | uh by looking at across region similarities |
---|
0:08:09 | so it's broken out into two step process |
---|
0:08:12 | uh the first is trying to |
---|
0:08:15 | assign appropriate plus one's uh |
---|
0:08:17 | plus one or minus one entries |
---|
0:08:19 | for the uh D matrices |
---|
0:08:21 | and it turns out that |
---|
0:08:23 | um |
---|
0:08:24 | and in in |
---|
0:08:25 | in a in expanding the problem |
---|
0:08:27 | this is um |
---|
0:08:29 | reduces to a quadratic binary optimization problem |
---|
0:08:32 | um |
---|
0:08:34 | and so we can use uh |
---|
0:08:36 | uh we we we can get an approximate solution with |
---|
0:08:39 | the |
---|
0:08:40 | max cut uh but but but by by using the algorithms |
---|
0:08:43 | to solve max cut |
---|
0:08:44 | uh the second step after we've got an R |
---|
0:08:48 | are D matrices |
---|
0:08:50 | wouldn't be to uh |
---|
0:08:52 | fine the or that call |
---|
0:08:53 | estimates |
---|
0:08:54 | uh with a with a greedy uh strategy |
---|
0:08:57 | and the way we do that |
---|
0:08:59 | is by selecting |
---|
0:09:01 | um elements |
---|
0:09:02 | that that are within groups and applying |
---|
0:09:05 | uh appropriate givens rotations to |
---|
0:09:07 | decrease the objective and so this is of a form of a |
---|
0:09:11 | block or set |
---|
0:09:16 | i to just a a a a a bit on a |
---|
0:09:18 | the second step um |
---|
0:09:21 | we can we can view the matrix C as a yeah |
---|
0:09:25 | as |
---|
0:09:26 | a noisy |
---|
0:09:28 | rotated version of |
---|
0:09:30 | of of a C R |
---|
0:09:32 | and |
---|
0:09:33 | and because the noise the the the eigen vectors |
---|
0:09:37 | uh of with a group similarity |
---|
0:09:40 | our berger |
---|
0:09:40 | so |
---|
0:09:41 | um |
---|
0:09:43 | the |
---|
0:09:44 | the uh factorisation we had the for uh of our initial estimate |
---|
0:09:47 | of |
---|
0:09:48 | are M equals |
---|
0:09:49 | the M M |
---|
0:09:50 | and D M V M R transpose |
---|
0:09:53 | uh |
---|
0:09:53 | we introduce |
---|
0:09:55 | another matrix you to get the M M U i'm the M V M M R transpose |
---|
0:09:59 | uh and |
---|
0:10:01 | and |
---|
0:10:01 | we view the U M and the D M |
---|
0:10:04 | as a a corrective measure |
---|
0:10:07 | to decrease the object |
---|
0:10:10 | so |
---|
0:10:10 | the D Ms |
---|
0:10:11 | as that's so so the use in the D uh are |
---|
0:10:15 | uh a there to incorporate |
---|
0:10:17 | the cross region similarities |
---|
0:10:19 | to correct the uh |
---|
0:10:21 | perturbed eigenvector |
---|
0:10:27 | so if we let uh |
---|
0:10:29 | you be D block diagonal |
---|
0:10:31 | uh |
---|
0:10:33 | version of of you want through you M |
---|
0:10:35 | then |
---|
0:10:36 | uh what we do is |
---|
0:10:38 | is is |
---|
0:10:39 | in our really approach is is you generate these givens rotations so we we pick |
---|
0:10:43 | um |
---|
0:10:45 | uh |
---|
0:10:45 | elements uh |
---|
0:10:46 | P T in Q T |
---|
0:10:48 | that belong the same group |
---|
0:10:49 | and we successively apply givens rotations |
---|
0:10:53 | to um at at at iteration C to decrease the object |
---|
0:10:58 | um um |
---|
0:10:59 | and at a a P change you T R |
---|
0:11:02 | are found |
---|
0:11:02 | uh using uh and and and and approximate greedy |
---|
0:11:05 | and and the angle is the sample the line search |
---|
0:11:11 | so |
---|
0:11:11 | uh look at some results |
---|
0:11:13 | uh |
---|
0:11:14 | so this was tested on on two data sets the first is set was the N nist handwritten digits |
---|
0:11:19 | um and second was the L B face day so |
---|
0:11:23 | for "'em" nist we have |
---|
0:11:24 | i ten groups |
---|
0:11:25 | for ten digits and for face database we have fifteen groups for |
---|
0:11:29 | uh a fifteen |
---|
0:11:30 | uh people |
---|
0:11:32 | so each group |
---|
0:11:33 | was |
---|
0:11:34 | uh transformed by by a random orthogonal matrix it's of M |
---|
0:11:38 | and then and and corrupted with noise |
---|
0:11:40 | so |
---|
0:11:41 | and you see before you |
---|
0:11:43 | the the the D block diagonal |
---|
0:11:46 | uh elements of this matrix are the with group similarity structures |
---|
0:11:50 | and the and the off diagonal |
---|
0:11:51 | elements are do you cross group |
---|
0:11:53 | and so you'll see uh |
---|
0:11:55 | you know |
---|
0:11:56 | the covariance of uh |
---|
0:11:57 | digit three and seven |
---|
0:11:59 | here |
---|
0:12:01 | point by the |
---|
0:12:05 | okay case at test accuracy we looked at |
---|
0:12:08 | um |
---|
0:12:10 | the |
---|
0:12:11 | our air estimate to noted uh |
---|
0:12:14 | by what what we're each |
---|
0:12:16 | we're reach block of |
---|
0:12:17 | of that rotation matrix is the dated |
---|
0:12:19 | uh |
---|
0:12:20 | and had uh we looked at |
---|
0:12:23 | the the inner product of um |
---|
0:12:26 | a have with a and uh and uh we took the average so so so this as the figure between |
---|
0:12:31 | minus one and one |
---|
0:12:33 | um |
---|
0:12:34 | and we see that uh |
---|
0:12:36 | that at iteration zero |
---|
0:12:39 | uh |
---|
0:12:40 | this is the result if we just take in |
---|
0:12:43 | to account the within group similarities and then |
---|
0:12:47 | as as a T increases the number of iterations increase |
---|
0:12:51 | uh out knowledge of across region similarities |
---|
0:12:54 | uh help us |
---|
0:12:56 | to teen uh a better match |
---|
0:12:59 | and uh we |
---|
0:13:00 | and this is true for both |
---|
0:13:02 | you know of the M nist |
---|
0:13:03 | and the old |
---|
0:13:07 | um |
---|
0:13:08 | going to the more adjusting problem of graph matching |
---|
0:13:11 | so |
---|
0:13:12 | uh |
---|
0:13:13 | each group in this case was was transformed by by some random permutation matrix |
---|
0:13:18 | and then perturbed |
---|
0:13:18 | with noise |
---|
0:13:19 | and uh |
---|
0:13:21 | the rather than solve for the |
---|
0:13:23 | uh am |
---|
0:13:25 | after after solving for each i M at at every time step you then |
---|
0:13:28 | use the hungarian algorithm |
---|
0:13:30 | to get the uh corresponding |
---|
0:13:33 | um permutation matrix |
---|
0:13:34 | and |
---|
0:13:35 | we can and and and we plot the accuracy |
---|
0:13:38 | uh with the correct number of uh |
---|
0:13:41 | permutation matches |
---|
0:13:43 | and uh |
---|
0:13:44 | for the M this there's there's a uh |
---|
0:13:47 | uh |
---|
0:13:47 | a a a a steeper improvement |
---|
0:13:49 | between only looking at within region two |
---|
0:13:53 | looking at across region |
---|
0:13:58 | a in conclusion oh O |
---|
0:14:00 | we pose the group |
---|
0:14:01 | uh version of the you side for this problem orthogonal cycle is problem |
---|
0:14:04 | uh we write an algorithm that's come to fit efficient given D givens rotations |
---|
0:14:09 | um |
---|
0:14:11 | and uh a simple to implement |
---|
0:14:13 | and |
---|
0:14:14 | there's results show that uh |
---|
0:14:16 | effectively you |
---|
0:14:18 | utilizing both within group but across group structure |
---|
0:14:21 | allows as to |
---|
0:14:22 | the estimation |
---|
0:14:23 | of the |
---|
0:14:24 | um and known |
---|
0:14:26 | orthogonal transformation |
---|
0:14:31 | thanks |
---|
0:14:34 | i |
---|
0:14:36 | sure |
---|
0:14:39 | yeah |
---|
0:14:40 | i i to los the obvious question is what can you say that convergence of the skein |
---|
0:14:49 | i would say that um |
---|
0:14:51 | that |
---|
0:14:52 | there are probably |
---|
0:14:54 | local minima |
---|
0:14:55 | given given |
---|
0:14:56 | given the constraints and so |
---|
0:14:59 | the key starting from that that |
---|
0:15:01 | that good initial |
---|
0:15:02 | and and the initial point |
---|
0:15:04 | so |
---|
0:15:05 | um |
---|
0:15:07 | more than that |
---|
0:15:08 | uh i can say |
---|
0:15:13 | no |
---|
0:15:14 | no |
---|