Transcript SVDslides

Smallest font
Please put away
electronic devices
Take out
clickers
Smallest font
Math
Tools
A representational view of matrices
Operation
D
a
t
a
D
a
t
a
Operation
Operation
Operation
D
a
t
a
D
a
t
a
Operation
Output
Operations
Input
A note on notation
•
•
•
•
m = number of rows, n = number of columns
“Number”: Number of vectors
“Size”: Dimensionality of each vector (elements)
[Operation matrix] [Input matrix]
• [number of operations x operation size] [size of input x number of inputs]
• [m1 x n1] [m2 x n2] = [m1 x n2]
• [2 x 3] [3 x 2] = [2 x 2]
• n1 and m2 (where the matrices touch) have to match –
size of operations has to match size of inputs.
• The resulting matrix will have dimensionality number of
operations x number of inputs.
Inner Product = Dot product
Dot product = Row Vector * Column Vector = Scalar
scalar
DP =
[1xn]
1234
5
6
7
8
[mx1]
[1x1]
1*5 + 2*6 + 3*7 + 4*8
(m=n)
70
Matrix multiplication
5 6
7 8
19 22
43 50
1 2
3 4
5 6
7 8
(1*5 + 2*7) (1*6 + 2*8)
(3*5 + 4*7) (3*6 + 4*8)
31 42
1 2
3 4
The outer product
• What if we do matrix multiplication, but when the
two matrices are a single column and row vector?
OP =
[mx1]
[1xn]
[mxn]
• Output is a *matrix*, not a scalar.
1
2
34
(1*3) (1*4)
(2*3) (2*4)
3 4
6 8
The transpose = flipping the matrix
1 2 3
1 4 7
4 5 6
2 5 8
7 8 9
3 6 9
NOT just a rotation!
Singular Value Decomposition (SVD)
• If you really want to understand what Eigenvalues
and Eigenvectors are…
• The culminating end point of a linear algebra
course.
• What we will use for the class (and beyond) to use
the power of linear algebra to do stuff we care
about.
Remember this?
Singular Value Decomposition (SVD)
M=US
Orthogonal
(“Rotate”)
T
V
Diagonal
(“Stretch”)
Orthogonal
(“Rotate”)
SVD
U
S
Output
Scaling
T
V
Input
A simple case
U
S
û1 û2
O
s1
s2
S
Outer
Product!
T
V
T
1
T
2
v̂
v̂
I
V= v̂1 v̂2
SVD can be interpreted as
• A sum of outer products!
• Decomposing the matrix into a sum of scaled
outer products.
• Key insight: The operations on respective
dimensions stay separate from each other, all
the way – through v, s and u.
• They are grouped, each operating on another
piece of the input.
Why does this perspective matter?
U
û1 û2
O
S
s1
s2
S
T
V
T
1
T
2
v̂
v̂
I
Is the SVD unique?
• Is there only one way to decompose a matrix M into
U, S and VT?
• Is there another set of orthogonal matrices
combined with a diagonal one that does the same
thing?
• The SVD is *not* unique.
• But for somewhat silly reasons.
• Ways in which it might not be unique?
Sign flipping!
U
û1 û2
U
û1 û2
S
s1
s2
S
-s1
T
V
T
1
T
2
v̂
v̂
T
V
-v̂
T
1
s2
v̂
T
2
Permutations!
U
û1 û2
U
û2 û1
S
s1
s2
S
s2
T
V
T
1
T
2
v̂
v̂
T
V
v̂
s1
T
2
T
1
v̂
Given that the SVD is *not* unique, how
can we make sure programs will all arrive
at the exact *same* SVD?
• Conventions!
• MATLAB picks the values in the main diagonal of
S to be positive – this also sets the signs in the
other two matrices.
• MATLAB orders them in decreasing magnitude,
with the largest one in the upper left.
• If it runs its algorithm and arrives at the wrong
order, it can always sort them (and does this
under the hood).
What happens if one of the entries in the
main diagonal of S is zero?
U
û1 û2
S
s1
s02
T
V
T
1
T
2
v̂
v̂
The nullspace
• The set of vectors that are mapped to zero.
• Information loss.
• Always comes about via multiplication by a zero in
the main diagonal of S.
• A one-dimensional black hole.
• A hydraulic press.
• “What happens to the inputs?”
• The part of the input space (V) that is lost
• Is a “subspace” of the input space.
What if S is not square?
U
û1 û2
S
100
010
T
V
T
1
T
2
T
3
v̂
v̂
v̂
What if the null space is multidimensional?
U
û1 û2
S
100
000
T
V
T
1
T
2
T
3
v̂
v̂
v̂
Nullspaces matter in cognition
• Categorization
• Attention
• Depth perception
The range space
•
•
•
•
•
Conceptually, a complement to the null space.
“What can we reach in the output space?”
“Where can we get to?”
Some things might be unachievable.
Where we can get to in the output space (U)?
How might we not reach a part of the
output space?
U
û1 û2
S
s1
0
T
V
T
1
T
2
v̂
v̂
The inverse question
• If I observe the outputs of a linear system and
watch what is coming, could I figure out what
the inputs were?
• Related problem: If you start with 2 things in
the input space and run them through the
system and compare the outputs, can we still
distinguish them as different?
• So when is the linear system invertible?
• How might it not be?
Inversion of matrices
• Matrix M can be inverted by doing the
operations of the SVD in inverse order.
• Peel them off one by one (last one done first):
• M = U S VT
T
T
• Inverting M:
• V S# UT U S VT
Q Q = QQ = I
I
I
The pseudoinverse
• “The best we can do”.
• We recover the information we can.
• The non-zero dimensions in the diagonal of S.
Next stop
• Eigenvectors
• Eigenvalues
• They will pop out as a special case of the SVD.
How else might we not reach a part of the
output space?
U
S
û1 û2 û3
10
01
00
T
V
T
1
T
2
v̂
v̂