Professional Documents
Culture Documents
Review of Le
ture 5
Di
hotomies
Break point
2
2.5
3.5
4
0.5
1
1.5
2
2.5
3
max
x1, ,xN X
|H(x1, , xN )|
Maximum # of di
hotomies
x1 x2 x3
Le ture 6:
Theory of Generalization
Outline
2/18
Bounding
mH(N )
To show:
mH(N ) is polynomial
We show:
mH(N ) a polynomial
Key quantity:
3/18
Re ursive bound on
B(N, k)
# of rows
B(N, k) = + 2
S1
S2+
S2
S2
x1 x2 . . . xN 1
+1 +1 . . .
+1
1 +1 . . .
+1
xN
+1
1
+1
1
+1
1
1
+1
1
1
...
...
...
...
1
1
+1
+1
1
+1
+1
+1
+1
1
+1
1
1
1
1
1
...
...
...
...
+1
1
+1
+1
+1
+1
1
1
+1 1 . . .
1 1 . . .
+1
1
1
1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
4/18
Estimating
and
# of rows
+ B(N 1, k)
S1
S2+
S2
S2
x1 x2 . . . xN 1
+1 +1 . . .
+1
1 +1 . . .
+1
xN
+1
1
+1
1
+1
1
1
+1
1
1
...
...
...
...
1
1
+1
+1
1
+1
+1
+1
+1
1
+1
1
1
1
1
1
...
...
...
...
+1
1
+1
+1
+1
+1
1
1
+1 1 . . .
1 1 . . .
+1
1
1
1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
5/18
Estimating
by itself
# of rows
B(N 1, k 1)
S1
S2+
S2
S2
x1 x2 . . . xN 1
+1 +1 . . .
+1
1 +1 . . .
+1
xN
+1
1
+1
1
+1
1
1
+1
1
1
...
...
...
...
1
1
+1
+1
1
+1
+1
+1
+1
1
+1
1
1
1
1
1
...
...
...
...
+1
1
+1
+1
+1
+1
1
1
+1 1 . . .
1 1 . . .
+1
1
1
1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
6/18
Putting it together
B(N, k) = + 2
# of rows
+ B(N 1, k)
S1
B(N 1, k 1)
S2+
B(N, k)
B(N 1, k) + B(N 1, k 1)
Learning From Data - Le
ture 6
S2
S2
x1 x2 . . . xN 1
+1 +1 . . .
+1
1 +1 . . .
+1
xN
+1
1
+1
1
+1
1
1
+1
1
1
...
...
...
...
1
1
+1
+1
1
+1
+1
+1
+1
1
+1
1
1
1
1
1
...
...
...
...
+1
1
+1
+1
+1
+1
1
1
+1 1 . . .
1 1 . . .
+1
1
1
1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
7/18
Numeri al omputation of
B(N, k)
bound
bottom
1
2
3
N 4
5
6
:
1
1
1
1
1
1
1
:
2
2
3
4
5
6
7
:
3
2
4
7
11
:
:
:
k
4
2
4
8
..
.
5
2
4
8
..
6
2
4
8
..
..
..
..
..
..
.
.
8/18
B(N, k)
bound
B(N, k)
k1
X
i=0
N
i
1. Boundary onditions:
easy
1
2
3
N 4
5
6
:
1
1
1
1
1
1
1
:
2
2
3
4
5
6
7
:
3
2
4
7
11
:
:
:
top
k
4
2
4
8
..
.
5
2
4
8
..
6
2
4
8
..
..
..
..
..
..
bottom
.
9/18
k1
X
i=0
N
i
=
=
=
=
k1
X
k2
X
N 1
N 1
+
?
i
i
i=0
i=0
k1
k1
X N 1
X N 1
+
1+
i1
i
i=1
i=1
k1
X N 1
N 1
1+
+
i
i1
i=1
k1
k1
X
X
N
N
X
=
1+
i
i
i=0
i=1
k1
N1
N
bottom
10/18
It is polynomial!
mH(N )
k1
X
N
i
|i=0 {z }
maximum power is
N k1
11/18
Three examples
k1
X
N
i=0
H is
positive rays:
H is
positive intervals:
H is
(break point k = 2)
mH(N ) = N + 1 N + 1
2D per eptrons:
(break point k = 3)
mH(N ) = 12 N 2 + 12 N + 1
(break point k = 4)
mH(N ) = ?
1 3
6N
1 2
2N
+ 12 N + 1
+ 56 N + 1
12/18
Outline
13/18
What we want
Instead of:
out
22N
We want:
22N
in
out
14/18
Pi torial proof
15/18
Hoeffding Inequality
Union Bound
VC Bound
space of
data sets
.
(a)
Learning From Data - Le
ture 6
(b)
(c)
16/18
What to do about
Eout
Hi
Eout(h)
Ein(h)
Eout(h)
Ein(h)
Ein(h)
Hi
17/18
Putting it together
Not quite:
2
2 N
P [ |E (g) E (g)| > ] 2 mH( N ) e
in
out
but rather:
out
1
8 2 N
e
18/18