subtitles.srt

1
00:00:00,200 --> 00:00:03,230
Hi everyone, I'm Patrick and in this video I show you how I would learn

2
00:00:03,262 --> 00:00:07,334
machine learning if I could start over. For context, I'm a machine learning developer

3
00:00:07,374 --> 00:00:10,910
advocate at AssemblyAI and before that I worked several years as

4
00:00:10,942 --> 00:00:14,854
software developer and ML engineer and I also teach Python and machine

5
00:00:14,894 --> 00:00:18,326
learning on my own YouTube channel. So I would say I'm pretty experienced

6
00:00:18,350 --> 00:00:21,750
in the field, but I know that the available courses out there

7
00:00:21,782 --> 00:00:25,078
can be overwhelming for beginners, so I hope to give you some guidance

8
00:00:25,126 --> 00:00:29,006
with this video. The demand for machine learning engineers is still increasing

9
00:00:29,110 --> 00:00:33,380
every year, so it's a great skill to have. I divided this learning path into

10
00:00:33,452 --> 00:00:36,924
seven steps that should take you about three months to finish. Of course,

11
00:00:36,964 --> 00:00:40,364
this can vary depending on how much time and effort you want to put into

12
00:00:40,404 --> 00:00:44,364
this, and I know that everyone learns differently or might have different goals.

13
00:00:44,484 --> 00:00:47,884
So this is just my personal take on how to learn machine

14
00:00:47,924 --> 00:00:51,436
learning. You can use this guide if you just want to explore machine learning as

15
00:00:51,460 --> 00:00:54,444
a hobby, but also if you plan to find a job in the field,

16
00:00:54,564 --> 00:00:57,900
I will mention a few more tips about the job search in the end.

17
00:00:58,012 --> 00:01:01,556
So lets jump into the study plan. The first thing I recommend is to lay

18
00:01:01,580 --> 00:01:04,804
the foundation with some math basics. Now you might say

19
00:01:04,844 --> 00:01:08,524
math is not really necessary anymore and this is partly true.

20
00:01:08,644 --> 00:01:12,004
The available machine learning frameworks abstract the math away

21
00:01:12,084 --> 00:01:15,588
and I know many machine learning engineers so dont need it in their day job

22
00:01:15,636 --> 00:01:19,116
at all. However, in my opinion, knowing the underlying math

23
00:01:19,180 --> 00:01:23,180
provides you with a better foundation and better understanding of how the algorithms

24
00:01:23,252 --> 00:01:26,996
work and it makes your life easier when you run into problems.

25
00:01:27,140 --> 00:01:30,212
Also, I think there is beauty in the underlying math

26
00:01:30,268 --> 00:01:33,940
that makes the machines learn. So for me, knowing the math

27
00:01:34,012 --> 00:01:37,844
sparked my excitement even more. Now you dont need to get too deep into

28
00:01:37,884 --> 00:01:41,636
this. A great website with free resources is Khan Academy.

29
00:01:41,740 --> 00:01:45,132
So my recommendation is just to take some basic courses and

30
00:01:45,148 --> 00:01:49,036
then move on. And then later when you do the actual machine learning course

31
00:01:49,100 --> 00:01:52,604
and dont understand everything, then come back here and learn the missing

32
00:01:52,644 --> 00:01:56,374
topics. Oh and by the way, you find all the resources and recommended courses

33
00:01:56,414 --> 00:01:59,766
in the description below. The next step is to learn Python. It is

34
00:01:59,790 --> 00:02:03,142
the number one programming language for machine learning and there is

35
00:02:03,198 --> 00:02:06,950
no way around it. All major machine learning frameworks are built

36
00:02:06,982 --> 00:02:10,718
with it and all major courses use Python for their exercises.

37
00:02:10,846 --> 00:02:14,110
So having decent Python skills is essential to build machine

38
00:02:14,142 --> 00:02:17,574
learning projects. Now you don't need to become an advanced software

39
00:02:17,614 --> 00:02:20,652
developer, but a little bit more than the beginner level would be great.

40
00:02:20,758 --> 00:02:24,112
One great thing about Python is that it is very beginner friendly,

41
00:02:24,168 --> 00:02:27,504
and in my opinion it's the best first programming language you can

42
00:02:27,544 --> 00:02:31,216
learn. I recommend two free courses on YouTube, one four

43
00:02:31,280 --> 00:02:34,392
hour beginner course and one six hour intermediate course,

44
00:02:34,448 --> 00:02:37,680
and then you should have a solid base. This step is what I call the

45
00:02:37,712 --> 00:02:41,792
machine learning tech stack and consists of the most important python libraries

46
00:02:41,848 --> 00:02:45,424
for machine learning, data science, and data visualization. This step

47
00:02:45,464 --> 00:02:48,896
is a bit optional because you can also pick up these skills later when you

48
00:02:48,920 --> 00:02:52,270
do the actual machine learning course, but I think it's great to build

49
00:02:52,302 --> 00:02:55,926
the foundation first and then it will be easier later. The three libraries

50
00:02:55,990 --> 00:02:59,262
I recommend at this point are numpy, which is the base for

51
00:02:59,318 --> 00:03:02,630
everything pandas, which is important for data handling,

52
00:03:02,702 --> 00:03:06,166
and matplotlib, which is needed for visualization. These libraries

53
00:03:06,230 --> 00:03:09,806
are used in almost every machine learning project. That's why I would include

54
00:03:09,870 --> 00:03:13,510
them in your learning path at this point. Again, you don't need

55
00:03:13,542 --> 00:03:16,598
to learn too much here. I recommend just following one

56
00:03:16,646 --> 00:03:20,000
free crash course for each library and then later pick

57
00:03:20,032 --> 00:03:23,552
up more advanced concepts if you need them. At this point you dont have

58
00:03:23,568 --> 00:03:27,488
to learn the machine learning courses like scikit learn, Tensorflow, or Pytorch.

59
00:03:27,576 --> 00:03:30,744
You could of course if you want, but these are included in the machine

60
00:03:30,784 --> 00:03:33,960
learning course I show you in a moment, so you can pick these up

61
00:03:33,992 --> 00:03:37,504
later. Now that weve covered the coding skills, it is finally time

62
00:03:37,544 --> 00:03:41,560
for the actual machine learning course. There are many great ones available,

63
00:03:41,672 --> 00:03:45,192
but the most popular, and in my opinion also one of the best

64
00:03:45,248 --> 00:03:48,680
ones, is the machine learning specialization by Andrew Ng on

65
00:03:48,712 --> 00:03:51,990
Coursera. This specialization includes its three courses.

66
00:03:52,062 --> 00:03:55,638
It got revamped just a few months ago and now includes Python

67
00:03:55,686 --> 00:03:58,766
with Numpy, scikit learn and Tensorflow for the code.

68
00:03:58,870 --> 00:04:02,422
So you not only learn all the essential machine learning concepts,

69
00:04:02,478 --> 00:04:05,758
but also get your first hands on experience with the ML

70
00:04:05,806 --> 00:04:09,182
libraries. It is extensive and it takes several weeks to finish,

71
00:04:09,238 --> 00:04:13,014
but it's worth it after these courses. I have one more recommendation for

72
00:04:13,054 --> 00:04:16,662
you. I suggest to implement a few algorithms from scratch

73
00:04:16,718 --> 00:04:20,056
in Python using only pure Python and numpy,

74
00:04:20,150 --> 00:04:24,092
for example by following my ML from scratch playlist here on YouTube.

75
00:04:24,148 --> 00:04:27,396
This is completely optional, but it helped me to properly understand

76
00:04:27,500 --> 00:04:31,340
some of the concepts from Andrew's course, and a lot of students have told

77
00:04:31,372 --> 00:04:34,660
me the same feedback, so check it out if you want to. Also, we plan

78
00:04:34,692 --> 00:04:38,180
to release an updated version of the ML from scratch course here

79
00:04:38,212 --> 00:04:41,396
on the AssemblyAI channel, so make sure to subscribe to

80
00:04:41,420 --> 00:04:45,108
our channel and don't miss it. Now I recommend getting even more hands on

81
00:04:45,156 --> 00:04:48,938
and learning more about data preparation. For this kaggle has

82
00:04:48,986 --> 00:04:52,066
awesome free courses on their website. I recommend at

83
00:04:52,090 --> 00:04:55,362
least the intro to ML and intermediate ML courses.

84
00:04:55,458 --> 00:04:58,826
They are lightweight compared to the previous one and some material is

85
00:04:58,850 --> 00:05:02,970
just a refresher for you, but youll learn more about data preprocessing

86
00:05:03,002 --> 00:05:06,202
and data preparation with pandas. Each lesson has a theory

87
00:05:06,258 --> 00:05:09,586
part and then some coding exercises. It also gives you a

88
00:05:09,610 --> 00:05:12,970
gentle introduction to the Kaggle platform and you learn how to make

89
00:05:13,002 --> 00:05:16,456
code submissions on Kaggle which is perfect for the next point.

90
00:05:16,560 --> 00:05:19,808
Now its time to practice as much as possible and

91
00:05:19,856 --> 00:05:23,096
apply your knowledge to real world machine learning problems.

92
00:05:23,200 --> 00:05:26,064
For this the best platform is kaggle.com.

93
00:05:26,184 --> 00:05:29,872
it provides thousands of different datasets and challenges where you

94
00:05:29,888 --> 00:05:33,240
can participate. Participating in challenges can motivate

95
00:05:33,272 --> 00:05:37,000
you a lot. Now I wouldnt try making it to the top or even winning

96
00:05:37,032 --> 00:05:40,976
prize money with this because to be honest, this requires true expertise

97
00:05:41,040 --> 00:05:44,828
and also a lot of GPU power. But I would still try

98
00:05:44,876 --> 00:05:48,132
to tweak your solutions multiple times by learning

99
00:05:48,188 --> 00:05:52,396
more about data preprocessing and also about hyperparameter tuning.

100
00:05:52,460 --> 00:05:55,820
You can then use Kaggle competitions to build your portfolio and put

101
00:05:55,852 --> 00:05:59,548
them on your cv. So in my opinion, Kaggle is an awesome

102
00:05:59,596 --> 00:06:03,412
platform and you should practice here as much as possible. Of course, you can also

103
00:06:03,468 --> 00:06:07,244
tackle other machine learning problems outside of Kaggle. It just makes your

104
00:06:07,284 --> 00:06:11,176
life a little bit easier because it provides you with the datasets, a platform

105
00:06:11,240 --> 00:06:14,736
to evaluate the projects, and there is a whole community around it.

106
00:06:14,800 --> 00:06:17,584
At this point you can already be super proud of yourself.

107
00:06:17,664 --> 00:06:21,144
And now in this last section, I want to give you a few more tips.

108
00:06:21,184 --> 00:06:24,864
If your goal is to get a job, the tasks of ML engineers vary

109
00:06:24,904 --> 00:06:27,952
a lot and it's not possible to know everything. For example,

110
00:06:28,008 --> 00:06:31,416
some positions are specialized in computer vision or NLP,

111
00:06:31,560 --> 00:06:35,448
or they require you to have experience with a specific ML framework

112
00:06:35,496 --> 00:06:38,978
or even mlops requirements like how to deploy and scale scale

113
00:06:39,026 --> 00:06:42,650
ML apps. MLops is a whole field on its own, so I may cover this

114
00:06:42,682 --> 00:06:46,306
in a separate video. My point is, you have to decide in which field you

115
00:06:46,330 --> 00:06:50,426
want to work and then look at the requirements in some corresponding job descriptions

116
00:06:50,490 --> 00:06:53,930
and then specialize in this direction. Another great tip I can

117
00:06:53,962 --> 00:06:57,754
give you, and this is something I wish I had done earlier in my career,

118
00:06:57,834 --> 00:07:01,546
is to start a blog. You can write tutorials, share what you've learned,

119
00:07:01,610 --> 00:07:05,066
which projects you have built, which problems you have faced along the way,

120
00:07:05,090 --> 00:07:08,442
and how you've solved them. By writing about a topic, you can deepen your

121
00:07:08,498 --> 00:07:11,984
knowledge and then you can use this as a resource on your cv.

122
00:07:12,104 --> 00:07:15,848
Trust me, this will increase your chances to get an interview a lot.

123
00:07:15,936 --> 00:07:19,216
Alright, that's my recommendation for a machine learning study guide.

124
00:07:19,280 --> 00:07:22,568
Again, this might not be suited for everyone. This is just how

125
00:07:22,616 --> 00:07:26,016
I would learn machine learning if I had to start over. Just one more

126
00:07:26,040 --> 00:07:29,608
quick addition. If you prefer learning with books, then you can check out these

127
00:07:29,656 --> 00:07:32,896
two books. Let me know in the comments if this was helpful,

128
00:07:32,960 --> 00:07:36,168
or if you have any other suggestions you would add to the plan.

129
00:07:36,296 --> 00:07:39,952
Don't forget to check out the resource list I put in the description below and

130
00:07:39,968 --> 00:07:41,808
then I hope to see you in the next video.