How to Make a Neural Network - Intro to Deep Learning #2.mp4.srt

0
00:00:00 --> 00:00:05
how do we learn how the times may change some Concepts stay the same 

1
00:00:05 --> 00:00:10
changing information Outlast the body it's stored in our brain but can be passed down 

2
00:00:10 --> 00:00:15
from generation to generation are brain is capable of synthesizing 

3
00:00:15 --> 00:00:20
what we call our five senses and from them creating a hierarchy of Concepts if we're 

4
00:00:20 --> 00:00:25
okay we can one attack while being supervised by a teacher directly while interacting with iron 

5
00:00:25 --> 00:00:30
and we can feel our surroundings see our obstacles and try to predict the next 

6
00:00:30 --> 00:00:35
if we try it first and fail that's okay through the process of trial and error 

7
00:00:35 --> 00:00:40
we can learn anything what is it that gives our brain to special capability 

8
00:00:40 --> 00:00:45
Nelson nature everything we've ever experienced or felt all our thoughts and memories are very 

9
00:00:45 --> 00:00:50
sense of self is produced by the brain at the molecular level are brain consists of a nest 

10
00:00:50 --> 00:00:55
a hundred billion nerve cells called neurons each neuron has three jobs 

11
00:00:55 --> 00:01:00


12
00:01:00 --> 00:01:05
whether or not the information should be passed on in the cell body or Soma and then get the sum of 

13
00:01:05 --> 00:01:10
channels passes a certain threshold send this results in Signal called the action potential 

14
00:01:10 --> 00:01:15
Ward's PA tax on to the next set of neurons 

15
00:01:15 --> 00:01:20
neural network in Python the rules that govern the brain's give rise to intelligence 

16
00:01:20 --> 00:01:25
it's the same album that invented modern language spaceflight Shia LaBeouf 

17
00:01:25 --> 00:01:30
what makes us us it's what's allowed us to survive and thrive on 

18
00:01:30 --> 00:01:35
but as far as we19
00:01:35 --> 00:01:40
existence there's the impending threat of climate change the possible 

20
00:01:40 --> 00:01:45
biochemical Warfare an asteroid impact 

21
00:01:45 --> 00:01:50
take our biological neural networks many generations to solve what if we could 

22
00:01:50 --> 00:01:55
is power on artificial neural network and had it run on a non Buy 

23
00:01:55 --> 00:02:00
magical substrate like Silicon we can give it more computing power and data and any 

24
00:02:00 --> 00:02:05
give me would be capable of handling and have it solved problems 8014 

25
00:02:05 --> 00:02:10
X Factor and we could alone in 1943 to early computer scientist 

26
00:02:10 --> 00:02:15
McCulloch and Walter Pitts invented the first computational model of a neuron 

27
00:02:15 --> 00:02:20
remodeled demonstrated a neuron that receive the binary inputs on them and if the Sun 

28
00:02:20 --> 00:02:25
secret a certain threshold value output of one if not I'll put a zero it was a 

29
00:02:25 --> 00:02:30
supermodel but in the early days of AI this was a big deal and got computer scientist talk 

30
00:02:30 --> 00:02:35
how about the possibilities later a psychologist in Brink Road 

31
00:02:35 --> 00:02:40
frustrated that include PetSmart 

32
00:02:40 --> 00:02:45


33
00:02:45 --> 00:02:50
the word for a single layer feed-forward neural network 

34
00:02:50 --> 00:02:55
cuz if they just closed in One Direction forward Incorporated 

35
00:02:55 --> 00:03:00
of weights on the inputs to give it some training set of input output examples 

36
00:03:00 --> 00:03:05
you should learn a function from it by increasing or decreasing the weights continuously 

37
00:03:05 --> 00:03:10
which example 

38
00:03:10 --> 00:03:15
to the input touch that after each iteration the output prediction gets more accurate 

39
00:03:15 --> 00:03:20
understand this process training 

40
00:03:20 --> 00:03:25
numpy as our dependency in our main function with first initial 

41
00:03:25 --> 00:03:30
Network which will later to find as its own class then print out it's starting weights for our 

42
00:03:30 --> 00:03:35
Prince one weekend we can now Define our dataset we've got four examples 

43
00:03:35 --> 00:03:40
Campbell has three input values can one output value for all zeros 

44
00:03:40 --> 00:03:45
function transposes The Matrix from horizontal to vertical 

45
00:03:45 --> 00:03:50
will train a neural network on these values so that given a new list of ones and zeros 

46
00:03:50 --> 00:03:55
predict whether or not the output unit 1 or 0 

47
00:03:55 --> 00:04:00
this is considered a classification task in machine Learning Network 

48
00:04:00 --> 00:04:05
using them as arguments to artrain function of the number 1049
00:04:05 --> 00:04:10
amount of times we like to iterate during training after it's done training will print out the 

50
00:04:10 --> 00:04:15
wait so we can compare them and finally will predict the output given a new input 

51
00:04:15 --> 00:04:20
our main function ready so let's not Define our neural network class when we initialize 

52
00:04:20 --> 00:04:25
the first thing I want to do is see diff will initialize our way 

53
00:04:25 --> 00:04:30
in a second and feeding them make sure that it generates the same numbers every time 

54
00:04:30 --> 00:04:35
this is useful for debugging later on random ways to A3 

55
00:04:35 --> 00:04:40
1 Matrix with values in the range of -1 to 1 with a mean of 0 

56
00:04:40 --> 00:04:45
are single neuron has three input connections and one output connection next right 

57
00:04:45 --> 00:04:50
activation function which it in our case will be a sigmoid it describes an s-shaped 

58
00:04:50 --> 00:04:55
we passed away did some of the inputs through it and it will convert them to a probability 

59
00:04:55 --> 00:05:00
between 0 and 1 this probability will help make our prediction 

60
00:05:00 --> 00:05:05


61
00:05:05 --> 00:05:10
are neurons to get to wait at some of our inputs will compute the. Product of our inputs and 

62
00:05:10 --> 00:05:15
weights this is how are weights govern the attention of how data flows in our neural-net 

63
00:05:15 --> 00:05:20
function will return our prediction now we can write out art rainbow which is the real meat of Arco 

64
00:05:20 --> 00:05:25
write a for Loop to enter a 1065
00:05:25 --> 00:05:30
function to pass the training set through the network and get the output value which is are for 

66
00:05:30 --> 00:05:35
next calculate the error which is a difference between the desired output and are predicted 

67
00:05:35 --> 00:05:40
what to minimize error as we train end up dating 

68
00:05:40 --> 00:05:45
calculate the necessary adjustment by Computing the dot product 

69
00:05:45 --> 00:05:50
and the error X the gradient of the sigmoid curve so less confident weights 

70
00:05:50 --> 00:05:55
traffic more and inputs that are zero don't cause changes to the weights this process 

71
00:05:55 --> 00:06:00
it's called gradient descent 

72
00:06:00 --> 00:06:05
calculates the derivative of our sigmoid which gives us its gradient or slope 

73
00:06:05 --> 00:06:10
Treasures how confident we are of the existing weight value and helps us update our prediction in 

74
00:06:10 --> 00:06:15
right direction finally once we have our adjustment will update our weights with that value 

75
00:06:15 --> 00:06:20
this process of propagating our error value back into our Network to adjust our weights 

76
00:06:20 --> 00:06:25
is called backpropagation what's the name of this baby in terminal because the training site is so 

77
00:06:25 --> 00:06:30
small took milliseconds to Trenton values updated themselves 

78
00:06:30 --> 00:06:35
federations Emily said it a novel input it predicted that the output was very light 

79
00:06:35 --> 00:06:40
we just made our first neural network from scratch anyways 

80
00:06:40 --> 00:06:45
propagation 

81
00:06:45 --> 00:06:50
navigate to update race update weight 

82
00:06:50 --> 00:06:55
navigate to update wait wait wait 

83
00:06:55 --> 00:07:00
Clinton knows map to ozan ones inputs go and add weights get some 

84
00:07:00 --> 00:07:05
that's that shitt to my sigmoid function get that heroin is real 

85
00:07:05 --> 00:07:10
that's what I use gradient descent and gives Direction and it doesn't 

86
00:07:10 --> 00:07:15
update weight and repeat 10000 * outputs or less 

87
00:07:15 --> 00:07:20
I'll be doing just fine so as dope as rosenblatt idea was in the decades following 

88
00:07:20 --> 00:07:25
give us any kind of noteworthy results only accomplished symbol 

89
00:07:25 --> 00:07:30
what is the world 

90
00:07:30 --> 00:07:35
is today we've seen an explosion in data and computing power 

91
00:07:35 --> 00:07:40
churches funded by the Canadian government held back to their belief in the power of neural networks 

92
00:07:40 --> 00:07:45


93
00:07:45 --> 00:07:50
or too many layers deep give it a huge dataset and lots of computing power 

94
00:07:50 --> 00:07:55
they discovered that you could outperform humans intact that we thought only we could do 

95
00:07:55 --> 00:08:00
this is profound are biological neural network is carbon based 

96
00:08:00 --> 00:08:05
chemicals like I see your colon glutamate and serotonin as signals 

97
00:08:05 --> 00:08:10
social network doesn't even exist in physical space it's an abstract concept 

98
00:08:10 --> 00:08:15
programmatically-created and is represented on Silicon transistors 

99
00:08:15 --> 00:08:20
difference in mediums they both develop a very similar mechanism for processing information 

100
00:08:20 --> 00:08:25
the results show that perhaps there's a lot of intelligence and coded into our universe 

101
00:08:25 --> 00:08:30
ever closer to finding it so to break it down at work is a bylaw 

102
00:08:30 --> 00:08:35
inspired algorithm that to identify patterns in data 

103
00:08:35 --> 00:08:40
what is a popular technique to train a neural network updating weights 

104
00:08:40 --> 00:08:45
descent retrain a mini layered a lot of data 

105
00:08:45 --> 00:08:50
lots of computing power we call this process deep learning 

106
00:08:50 --> 00:08:55
is Pluto made a really slick IPython notebook 2 demo 

107
00:08:55 --> 00:09:00
3D regression as well on a climate change data set other week 

108
00:09:00 --> 00:09:05
runner-up is a man who can play the bonus with great results the challenge 

109
00:09:05 --> 00:09:10
it's a great I'm not one not two but three layer feed forward neural network 

110
00:09:10 --> 00:09:15
just jump I post a comment in the winter in one week 

111
00:09:15 --> 00:09:20
subscribe and for now I got to update my weight so thanks for what