yeah

a risk is switch to the second

present a sent in this in this you in this a we we you present the fast they might those

new with them

yes

and is

this is a but you on what of well with a bit that's caff be and i C okay

and uh

oh

so a

i you already know that the um

this

yeah are another

kind of but with some for at the end and for C be that

the the on a the on S one egg with um

and one as the only one in some these

yeah the dam goes you than it them

but usually for them most new data can you we we on we need to compute the that be beyond

oh the gradient and the hessian and and usually do they has and very very few already last

and

uh it basically that's

we

was up we need to in but very hot

but a at that possibly has an and in in this

in this present is that we we we to do how to

avoid this problem

how to how to avoid

how to

a lot but enough that has an and all so how to a the

the lower the last

in but up last

has yeah

and now is the a life

oh

might talk that we we you

resent deeply side it press for the good

the copy of method

and we will find the

the fast way to compare

to compute looks

the gradient

i also to

compute a process as N

and it here we we

with an

the low rain and that's one for the a possibly has and

and then by on this

but you by on this low when it just when we we fight

we we i

we we we present

a very it's so that need to

to avoid

the input of last has

i skip this require already meant something

for

reverse light

okay

a you know the to derive with that

the the atlas electrics um we can see that the same the similar conversant robot whereabouts roy

in nineteen seventy nine

but here you see here this is a conference on

two we to approximate the ten so by

but extract the

with

this the two

to write really says that um is the

okay this term

is rest and it's right here and this time we present here

this time is to inform the nonnegative trends on the front end

i N

and this time to an problems

spice you can train

and the car far and bit eyes

zero are uh plus they've

and is a see here

to that

the on as one of the lead that we have best

all that on the fact that i the same time

and he'll to be two two we before all this we use this

but that's room

with

with i we've fit i is that that's

okay you C

this the best the is a set up fact the i one

and then you

code for i on the on the but is a seven

to very long for that

but long list that and we employ this

a back room

do you best

the fact

but

you see here is we we need to compute the the copy and here i'm node that has sent the

possibly may has sent here

and also the gradient here

and the

a call we the also we of go we need to need to compute you

the the cool and

from the be and we compute the gradient

for a for in this fall

and also so be that has an approach me has an

for in this fall

and

a probably C come men

i i a high computational cost

and of course the of this but might is side

you one here

and a lot of time and also

yes

and

we we we present here than

the way to avoid this problem

it this line we'll

so that how to

compute the will be and matters

this do though it the bustle

the river a of the approximate ten so

which search but only one come but

and

we saw this this

we be is spread in the very

come come

vol

with

this one is that correct up

correct upper up all the

common a

in in on the fact that it's that the the fact that and

and B and is that the rotation matrix

it better

okay to say here this this breath the relation between

the vector is a stand up

okay so

and vector is a

and vector is a set of the most and

my matrices a send up to time so

quite complicated here

you see if you were to i don't the data and

and this one leader

you where the wide the mode and

methods and of the ten so

and this italy's relation between

two batteries and

and we fine

the compact form for that the the copy be and map so here

and with few we we from this result we

find the weighted method

this one is the uh

the final

result for the gradient vectors

and it with you see here this one is the gradient of the tensor with respect to the fact though

and

the fact that one

and this one is the right

the great and

of the

a cross and tensor

which are but to the fact the and

and the got my mike this is press here the that the how to map product up

on

this time

what is that the fact the and

is that the the something

and of some this this crime matter

at the nest

slide

we we present

how to compute the

approximate

yes

the idea is this we we've split

we can see that the the possibly has and that though

and by and

block might be

and we we fight the plug i that is the press for for every block here

is

this dist

to and you with

how to you

that was sub just

has us a but just

like did you have a you

if and by it

for

say this is them

the diagonal block this

is and equal to M you you have to um the diagonal plot mattress and C

compute by

yeah my

and

i am and go product up a got my

with the and it it matters

a a um of sci ice ice of N

and is it that the diagonal vector

and for that off diagonal my

for the up

diagonal diagonal method

when

and you know in a equal to N

we we we can decompose

this

supp might in into

this four

this it is a diagonal matrix

this the that block diagonal matrix enough for

i i know what it

is it problem

i i matter and is also the block diagonal matter

we be spread here

in the next line

so

a very in i property of that of possibly has an that that we can

it's spread

the approximate has an and that the low end

and that one

you see year

so i say yeah that a i all method

we we already defined though

re slide

and is that the point

but but and method

and

see here here that the right block low matter

or within this fall

i K yeah A the

square matrix

oh side and

and ask way by and ask where

and this well

in the is the mac of the K method here even in this fall

be here

okay

ah

it in S like we you see here how to

a which to like the case and

and how

how we can

and to spend a little way that's one

i just one for the prosody has and yeah

yeah

the the has the a to my has an

and it is a metric

see you the plot like no mapping

see you that the also plot that all method

and case have where

particular

such that

and of car yeah if we employ

deep if we employ the by no

binomial

it but to your we can it

we can but we can of one

the in but a blast scale

a little the last has in method

by this

for

and if we compute here is that a compute the in but uh

the approximate

has and we

all only compute it looked at the you might rate the came at rate and also in lot of this

style

and this time have very

small

the by some because this is this term is

has i'll up and by K and

and K

and a

and a square by and ask square

and in but

the same her

we can quickly compute

by

a because it brought no matrix

so is slide up problem

is that of compute the in of the whole that the whole take we compute in but it's

sub method along the diagonal

so do you have here

and we also had how to compute the in but at that a least

is have very is so uh

nine

it press here

and finally from that them "'cause" new done of that room we fight with that

we we formulate the fast them let's be learned with some that to we reply here

i the approximation has has

and we define you

this

the fast them book use like

which

the with uh okay this

this term

this term we that

no this um

i you see

we so

this them with that this this

this a

and

and the L make

it

this border

with uh this the

okay also it's spread by the

block diagonal i dunno method

so finally

you see here the the the true this

most one

with the of the

the the five but to the much smaller than the this the approximation

uh has and then and these K we can we do the at the clear do the commit

so the of the

in a of this method

in the very don't to the in but that this

the possibly has and

and a call we don't need to control the to go beyond method

we not only to compute the

a possibly has an

maybe a nice

okay okay

so in this snow the next slide you see here

how to to the power with a how do to the the but but that in in the

you know an increase um

a very simple it and also very if if you soon that's a proposal were about rowing two thousand nine

that's of and

you nine

uh that's but more like that it we it in line

on the

the low power by real met um but power what's so by the la

last

enough a value and then we greater only

we do

you do the bar with and then we can

after the fire ten iteration we can

yeah

the can was and that the that's solution

and

we

we don't use this approach

we you are not a

and this you this one T by on the cake Q T condition

and this list to

the medium i of this problem

this for that's of that too

this condition

okay

and as to the simulation

is assume a some place and you see here for three dimensional tensor a with side what one and right

by one hundred by one hundred combo by ten component

yeah

but

the common

we for to be collinear we the auto by this form

and we see here

with the

multiplicative kl K with them what what deeply get the L

the green light

and that the i

is from a up the

or

one one right it ways that no one that fred

i not yet and

one ten held and B

that's how the in twice and then the and with only

get a convert to the solution and

and i like that on the L

the fast lm level and is them with quickly can what up the only

one at right right

but not case and not a that it we one that we

test with one tell than by one column

i want how than i with the common

you to here one attract recommended

it's really he'll

and the

fast and where make on is also quickly

fit the data up the

fifteen no sitting it weights and

i and all right with some can of fit the data

i is to to much but maybe i skip this like

is used for to this prior to

for of the method but maybe there with

and

or

this you know uh an application with the leave at the faded the way that we

we classified the data

that's we from the data that we have forty four interest for a and that we put from the fight

label

by the cable with flat it to this just kind of ten so C to by sitting

by

that T two to T two because we have yeah i

i orientation and for scale

and of call we have yet the last of and is that number of plays

for right

and

we

it truck of from this

ten so

for the first

first we track by

thirty two

and T F that we have here

and the S like that we employ an an map to talk dove

three too

from the last

one

the life factor

and the but form we

of and here you see

but the K and it with on is only

i was N

with the us

i T i percent

is

maybe maybe though with the

are you less

with night night it's three was and but our and with need

not for was a and if we employ this was can trying the ten i

one as that was sent with ten

face and all is not enough ten fate take class

that's me we have one and right

for

and we

it to the good clothes and

a you see here we already but both a fast and with them too

for for et and a map the fast that was knew that for N

and the M

and that was don't we

a you don't

we to compare you are the last scale that will be and and also that scale

has has C and

the process may has them

and we avoid

they vote of the the last

has

and it

and you are virtually you see here would have the a into trend

the fast them was new than

we we the fast lemme i rammed

and we term for canonical only

um only to decompose and

that these these two D dizzy

similar to the one the one of the wooden and we can where you but row and

and the mostly but

oh and we is further

thank you

i

uh yeah think to the

okay we we we we may the the net and also is similar to we man that the relative error

two

you see here

this is the right of the arrow

but we we do is read how to compute this error of this is very simple that's we

we

okay

from the conference and you

this so we compute this error and divide by norm up the so

yeah

this your L T you ever

yes

no i i you these

for from one is known so we don't you are not the kind of um is a one

i