A Novel Narrative on Neural Networks
@captainsafia
safia.rocks
http://j.mp/nnn-safia
Follow along!
What's up, Doc?
What's deep learning?
What are neural networks?
How do neural networks work?
When should I use neural networks?
Applying neural networks to a character recognition problem.
Wild, Wild, West!
What's deep learning?
A way for machines to understand the world.
AI
➡
Machine Learning
➡
Representation Learning
➡
Deep Learning
What are neural networks?
A Brief History of Neural Networks
Warren McCulloch and Walter Pitts
MADLINE and ADALINE
1959
First multilayered neural network
1975
John Hopfield, 1982
Neural Networks Win!
2006 to 2012
How do they work?
How likely is it that someone will get a job at Google?
Features
Years of programming experience, college GPA, number of stars on most popular GitHub project
Label
Percentage of "Yes" votes from hiring committee
years_exp
gpa
stars
percent_yes
4
3.4
2
60
3
3.1
12
80
7
2.9
23
46
1
3.8
243
77
4
3.2
1000
91
4
4.0
0
30
Activation Functions
Transform an input to a different domain in the output
Introduce non-linearity to the network so that our results are not just a linear combination of the weights
Sigmoid function
W1
I
H1
W2
O
W1
I
H1
W2
O
W1
I
H1
W2
O
W1
I
H1
W2
O
W1
I
H1
W2
O
Cost Function
MSE = \frac{1}{n} \sum ({X}' - X)^2
M
S
E
=
1
n
∑
(
X
′
−
X
)
2
MSE = \frac{1}{n} \sum ({X}' - X)^2
M
S
E
=
n
1
∑
(
X
′
−
X
)
2
MSE = \frac{1}{n} \sum ({X}' - X)^2
M
S
E
=
1
n
∑
(
X
′
−
X
)
2
MSE = \frac{1}{n} \sum ({X}' - X)^2
M
S
E
=
n
1
∑
(
X
′
−
X
)
2
E(W1, W2) = \frac{1}{n} \sum ({X}' - X(W1, W2))^2
E
(
W
1
,
W
2
)
=
1
n
∑
(
X
′
−
X
(
W
1
,
W
2
)
)
2
E(W1, W2) = \frac{1}{n} \sum ({X}' - X(W1, W2))^2
E
(
W
1
,
W
2
)
=
n
1
∑
(
X
′
−
X
(
W
1
,
W
2
)
)
2
\frac{d E(W1, W2) }{d W1, W2}
d
E
(
W
1
,
W
2
)
d
W
1
,
W
2
\frac{d E(W1, W2) }{d W1, W2}
d
W
1
,
W
2
d
E
(
W
1
,
W
2
)
W1
W2
E(W1,W2)
Multilayer Perceptron for Classification
When should we use neural networks?
Can be used for classification or regression problems
Comfortable with a complex model representation
Comfortable with doing a lot of data normalization
https://git.io/vr4qe
Let's get our hands dirty!
Made with Slides.com