1 Video View·Dec 3, 2023
Want to learn Machine Learning (ML), Deep Learning (DL), 5G, 6G, and PYTHON?
Check out our schools below!
https://www.iitk.ac.in/mwn/MLDL/index.html
https://www.iitk.ac.in/mwn/IITK6G/index.html
IIT Kanpur Intensive Training School (ITS) on PYTHON for Machine Learning, Neural Networks and Deep Learning
2nd to 22nd December 2023
https://www.iitk.ac.in/mwn/MLDL/index.html
Welcome to the IIT Kanpur Intensive Training School (ITS) on PYTHON for Machine Learning, Neural Networks and Deep Learning. This school, spanning over 3 weeks, is an intensive program for scholars, students, faculty members, industry professionals and R&D staff aspiring to get an in-depth exposure and learn hands-on implementation of the cutting edge algorithms and software for Machine Learning (ML), Neural Networks (NN) and Deep Learning (DL) toward projects, research and placements to further their career. The school will include extensive assisted PYTHON programming sessions, and numerous important packages such as NUMPY, LINALG, MATPLOTLIB, SCIKIT-LEARN, TENSORFLOW, KERAS. The school also includes problem solving sessions to prepare for tests/ job interviews in ML and DL. For an extensive list of lecture modules and hands-on programming projects, please check the PROGRAM
IIT Kanpur PYTHON + MATLAB Training School on 5G/ 6G Wireless Technologies
2nd to 22nd March 2024
https://www.iitk.ac.in/mwn/IITK6G/index.html
Welcome to the PYTHON + MATLAB Training School on 5G/ 6G Wireless Technologies: MIMO, massive MIMO, mmWave MIMO, NOMA, OTFS and IRS. This is a unique cutting edge project based training school that will feature exhaustive lecture modules and several PYTHON/ MATLAB projects/ case-studies on the latest MIMO, Massive MIMO, mmWave MIMO, NOMA technologies for 5G. Also, with 5G systems being rapidly deployed around the globe, there is accelerated development of Beyond 5G (B5G) and 6G wireless technologies that break the barriers of 5G. OTFS is a cutting-edge 2D modulation technique that operates in the Delay-Doppler domain, while Intelligent reflecting surface (IRS) is another novel technology that enables dynamically altering the radio propagation environment to achieve a phenomenal boost in the capacities. Together, these dual technologies are expected to power the shift to 6G. This training school will feature exhaustive tutorial style lecture modules and several intensive supervised PYTHON/ MATLAB/ OCTAVE programming modules so that participants are able to gain practical hands-on experience of working on state-of-the-art 5G/ 6G projects
All modules will be held on evenings and weekends for the convenience of participants. The extensive projects and PYTHON programming for MACHINE LEARNING (ML), NEURAL NETWORKS (NN), DEEP LEARNING (DL), 5G and 6G are of significant value to participants of all backgrounds.
How do these programs benefit YOU?
UG/ PG students: Learn the latest programming techniques in PYTHON and various packages such as NUMPY, LINALG, MATPLOTLIB, PANDAS, SCIKIT-LEARN, SEABORN, TENSORFLOW, KERAS together with practical ML/ NN / DL/ 5G/ 6G knowledge for projects/ thesis and also to gain an unbeatable edge in placements!
PhD Scholars/ Faculty members: Use PYTHON, and various packages such as NUMPY, LINALG, MATPLOTLIB, PANDAS, SCIKIT-LEARN, SEABORN, TENSORFLOW, KERAS for research and also to establish virtual labs or for project guidance in ML/ NN / DL/ 5G/ 6G Technologies with real world datasets!
Industry Professionals: Take your skills to the next level by learning PYTHON, and various packages such as NUMPY, LINALG, MATPLOTLIB, PANDAS, SCIKITLEARN, SEABORN, TENSORFLOW, KERAS, together with principles of Module Design and Analysis using Practical Data Sets for ML/ NN / DL/ 5G/ 6G Technologies!
Learn the latest ML/ NN/ DL/ 5G/ 6G Algorithms, PYTHON Programming & Libraries from experts at IIT Kanpur
PYTHON, Machine Learning, Neural Networks, Deep Learning, 5G and 6G are the top skills for placements in leading global MNCs such as Google, Amazon, facebook, Microsoft, Qualcomm, Broadcom, Samsung, Mediatek and many others!
Our industry participants have come from top global MNCs, Indian R&D Labs below and many more!
Our academic participants have come from IITs, NITs, top national and international educational institutions and more!
Contact Us:
e-mail: [email protected]
Join/ follow us on our social media sites
https://www.facebook.com/people/MWN-Lab-IIT-Kanpur/100063786175713/
https://twitter.com/LabMwn
https://www.linkedin.com/in/mwn-lab-7236231ba/?originalSubdomain=in
[Music]
[Music]
hello welcome to another module in this
massive open online course so we are
looking at same as DP there is
semi-definite programming and its
application in the context of my mode
detection that is how to reduce the
complexity of the MIMO detector that
there is my maximum likelihood detective
okay so we are looking at is TP for my
ML detection okay
and as DP employs an interesting
constraint if you remember this this is
a positive semi-definite constraint that
is but basically this matrix has to be
this linear combination of matrices has
to be positive so positive semi definite
this is termed as a linear matrix
inequality LMI okay so SDP is
interesting in the sense that it employs
a matrix inequality
this is a linear matrix inequality SM as
DP and force has a linear matrix
inequality that is what is novel about
STP and we are specifically looking at
this my mode detection problem where we
have a received vector Y bar minus H is
the MIMO channel matrix H X bar square
and we have to perform our mean that is
minimized that is fine the X bar
belonging to this set s which minimizes
y bar minus SX bar square and the
problem with this is if you look at this
set the cardinality of the size of that
set is basically M raised to T where m
is the number of symbols in the
consideration okay
and of course T this is equal to the
number of transmitted
t equals number of transmit antennas and
so basically this is exponential in the
number of transmit antennas which is a
very high complexity okay
so this is exponential in number of
transmit antennas
so how can SDP help in this context for
his DP what we do is let us simplify the
cost function this is y bar minus HX per
square this is your objective objective
function or cost that has to be
minimized okay and I can write this as
follows y bar minus HX bar this is norm
square of a vector so I could write this
as vector transpose times itself okay so
I can write this proceed by writing this
as Y bar minus HX bar transpose Y bar
minus HX bar the simplification we have
seen previously also so I'll go fast
over this y bar minus y bar transpose
minus X transpose H transpose into y bar
minus H X bar which is equal to Y bar
transpose Y bar minus X bar transpose H
transpose Y bar minus y bar transpose H
X bar plus X bar transpose H transpose H
transpose X bar okay so this is
basically the cost function simplified
customs now here I am going to do
something interesting I am going to
write this as follows and convince
yourself this can be written as the
following thing okay this can be written
as equivalently as X bar transpose one
so I'm making a column vector so X bar I
am stacking it along with this number
one so I'm making it T plus one
dimensional column vector okay X bar
transpose one and this matrix I can
write this as H transpose H minus H
transpose Y bar minus y bar transpose H
why bar transpose Y bar times X per one
okay and if I call this T plus 1
dimensional vector no s bar or rather is
bar transpose this I call as as well
this I can call as the matrix L okay and
therefore what I have is I have this
vector s bar which I am defining as X
bar 1 this is a t dimensional
and this is so this is single element so
this net vector will be this vector will
be T plus one dimensional and if you
look at this vector this is a matrix L
which is given as follows so L is the
matrix which is given as follows this is
H transpose H minus H transpose Y bar
minus y transpose H and this is y bar
transpose Y bar okay so therefore we
have this matrix l and therefore now I
can write this basically if you look at
this this is s bar transpose L into s
bar so we have written this as is par
transpose and into s bar so I have
recast this as weighted norm
minimization problem that is we have
written this in the form of s bar
transpose L times s bar where L can be
thought of as a weighting matrix okay
and now I can equivalently minimize this
minimize so I can minimize this s bar
transpose L s bar okay so I can minimize
s bar transpose L s bar however the only
problem here is that of course we have
constraints now what are the constraints
now remember if you look at s bar still
we have I mean the problem we have just
recast it the problem has not changed we
still have if you look at s bar which is
X bar comma 1 so if you look at s bar so
if you look at the ayat element okay so
your for 1 less than equal to I left so
if you look at the ayat element that is
si for 1 less than equal to I less than
equal to t that still has
to belong to the constellation so let us
say this is BPSK constellation so this
has to be plus or minus one okay so
these elements that is the first P
elements are still the symbols that is X
1 X 2 up to X T so remember these still
have to belong to the constellation that
is each of this can only take two put
two values that is plus or minus one
which is where the problem is arising in
the first place all right because you
have this large number of factors over
which you have to search now I am going
to manipulate this cost function in a
step by step first note that each
magnitude si square in fact for the t
plus 1 element is also 1 so all
magnitude si square equals unity or in
fact in this case simply considering
real symbols you can note that simply si
square equals unity okay so if you look
at s Bar s bar transpose which is
basically s 1 s 2 up to s T plus 1 into
its transpose s 1 up to s T plus 1 if
you look at the diagonal elements of
this so this is not a diagonal matrix
but if you look at the diagonal elements
all the diagonal elements will be there
is there are also off diagonal these are
the of
diagonal but if you look at the diagonal
entries all the diagonal entries are
equal to one because each si square is 1
therefore all the diagonal elements of
this as far as what transpose are what
our unit but now what I am going to do
is basically not you realize the first
thing as we have shown these two are
equivalent so minimize nor Y bar minus H
X bar square can be equivalent ly
written as minimize s bar transpose L s
bar okay
each si belongs to still this symbol of
plus or minus one assuming BPSK binary
phase shift key and for of course one
less than equal to I less than equal to
P now what I am going to do here is now
I am going to employ this thing so now
first step one observe that s bar
transpose LS bar this is a scalar
quantity this is a number so I can write
this as trace because if you take the
trace of a number it remember trace is
the sum of the diagonal elements for a
square matrix so a single number is a
special case of square matrix so the
trace will yield the number itself so
this I can write this as trace s bar
transpose LS bar now using the property
trace a B equals tree is B a so this
becomes trace of L s bar s bar transpose
which I can now write as trace L types
where what is s s equals the matrix as
far as bar transpose which is of course
positive semi-definite because I can
express this as s bar as bar transpose
and diagonal of s equals basically you
have all the elements or unity so this
has to be this vector 1 bar and further
we have seen that s is positive
semi-definite s is greater than equal to
2 so now I can equally formulate this as
so equivalent problem will become
minimize equality problem is minimize s
bar transpose LS bar subject to the
constraint or in fact not minimized s
bar transpose LS / I apologize this is
minimize trace of LS subject to the
constraint that s is a positive semi
definite matrix that is s is greater
than or equal to 0 first let us write
the diagonal so diagonal of s equal to
unity or not unity that is each element
of the Dom the diagonal of s is 1 so
diagonal of s is 1 bar the vector of
one's s is a positive semi definite
matrix
ok so s is positive definite semi
definite further s can be written as s
bar s bar transpose now of all the
constraints this is the most difficult
most difficult non convex constraint
because it is non convex this is known
as a Rank 1 constraint
this is known as the rank one constraint
because s equal to s bar is that a
single vector expressed as a vector s
bar s bar transpose so Manor you have a
vector matrix of this form which is
vector times its transpose then it will
basically rank one constraint and this
is a non convex constraint so ensuring
this is very difficult of constraining
this optimization problem using this
rank one constraint is benefit so what
we do is naturally what we do when we
encounter something difficult or
difficult to impose we simplify in this
case we simply ignore this this is known
as an SDP relaxation okay so we relax it
so this rank 1 constraint makes it non
convex so it makes it no less DP so we
relax it as an SDP that is ignore this
rank 1 constraint so what we do ignore
the rank 1 constraint so we ignore the
rank 1 constraint
obtain sdp this is term as
this is termed as SDP relaxation this is
termed as STP relaxation okay so what we
have now is an interesting problem that
is we can equivalently so our ml decoder
can be equivalently written as ml my
mode decoder for that matter can be
equivalently written as minimize trace
of LS subject to the fact that diagonal
s equals the matrix of one's s is
positive semi-definite and that's it
okay
and this yields an approximate solution
so this is significantly lower
complexity so now from exponential
complexity we have come to we are convex
no this is a convex optimization problem
okay now this problem is convex in fact
this is an SDP it has a linear matrix
inequality so this is an SDP convex this
is an SDP
now while the original problem is
exponential this problem has much lower
complexity
this has much lower complexity and
therefore it's very amenable to
implement this in practice very suitable
for practical implementation the only
thing is it yields an approximate
solution of course it's solution that is
actually shown to be very close to the
optimal eval solution so this yields
approximate solution
close to the ml solution now once you
find s
now once you find as how to find s or
how to find s bar rather so the point is
you find this matrix s is equal to s bar
s bar transpose but not guaranteed
because we have ignored the rank 1
constraint not s is not guaranteed to be
a rank one matrix S is not guaranteed to
be of the form s bar as per transpose so
now since we ignored it s is not
guaranteed
so s is not guaranteed so in this
context how do you find s1 is very
simple answer is where if you find the
best rank one approximation to s --t
that is the key here Phi into this rank
one approximation
find best ranked one approximation to s4
that implies we use the eigenvalue
decomposition of s and this is also very
interesting how do you find the best
rank one decomposition of a matrix
square matrix or symmetric matrix use
the eigenvalue use the eigenvalue
decomposition remember the eigenvalue
decomposition is as follows i can write
s as Q lambda Q transpose where Q is the
matrix of eigenvectors
and this is diagonal matrix of
eigenvalues
this is the diagonal matrix of
eigenvalues and then I can write this as
summation I equals 1 to n n plus 1
lambda I Qi bar Qi bar transpose where
lambda I equals ayat eigenvalue and Qi
bar equals the ayat eigenvector and
therefore this is simply nothing but
lambda 1 Q 1 bar Q 1 bar transpose plus
lambda 2 Q 2 bar Q 2 bar transpose plus
lambda 3 Q 3 bar 2 3 bar transpose now
since this is positive semi-definite
since s is PSD this is PSD note that
lambdas are greater than or equal to so
this implies lambda is greater than
equal to 0 so I can always arrange them
in decreasing order all right so this is
lambda is a real quantity is in fact
they are greater than equal to 0 so I
can always arrange lambda 1 greater than
equal to lambda 2 greater equal to
lambda 3 so on so so let lambda 1
greater than equal to lambda 2 greater
than equal to lambda 3 and so on so
implies lambda R 1 equals maximum eigen
value then the best rank one
approximation
best-ranked one approximation is simply
this that is you choose s equal to
lambda1 so ignore the rest okay just
simply choose s equal to the largest
eigenvalue eigenvector outer product of
eigenvector corresponding are just agree
so best ranked one approximation is
simply s is approximately equal to
lambda 1 q 1 bar u 1 bar transpose which
i can always write a square root lambda
1 q 1 bar square root lambda 1 u 1 bar
transpose which is now equal to err s
hat SI transpose au s equal to or rather
s hat equal to s hat as a transpose and
the final step is is equal to remember
your vector X bar comma 1 so that
implies s hat equals x hat comma 1 ok so
by choosing the first T symbols of aside
you get transmitted symbols X in X bar
okay so by choosing
one can estimate or detect rather the
transmitted
symbols X and from that of course you
can get the transmit vector X that is
how you basically so you take the
original ML decoder right recast it in a
different form all right and then you
relax the critical thing here is that
relax the rank one constraint that makes
it a semi definite program this process
known as SDP relaxation from the SDP
relaxation you get s which is a positive
semi-definite matrix from that you
perform the eigenvalue decomposition but
get the best rank when approximations
that will be in fact there is nothing
but the principal eigenvector
alright the scale to principle I can
given in terms of the scaled principal
eigenvector of s from that you take the
top T symbols those are your MLS alright
so let us stop here and continue the
subsequent words thank you very much
[Music]
[Music]