00:00:00
look at all this math and code you need
00:00:01
to understand to learn machine learning
00:00:04
it can be very hard for me at least it
00:00:06
was as well until I learned these five
00:00:09
Secrets which honestly aren't even
00:00:11
Secrets but no one really teaches you
00:00:14
them although everyone should know them
00:00:16
I mean I spent the last 3 and a half
00:00:19
years studying machine learning and it
00:00:21
took me way too long to learn these
00:00:23
secrets on my own so let me reveal them
00:00:25
to you so that you don't have to
00:00:27
struggle for that
00:00:28
long
00:00:32
you're thinking of math the wrong way
00:00:34
around which already sets you up for
00:00:36
failure at the very beginning but isn't
00:00:38
your fault back when I started learning
00:00:40
machine learning some of my professors
00:00:43
would simply throw a formula on screen
00:00:45
and tell us this is the loss function
00:00:47
for a decision tree and that was it I
00:00:50
and most other peers were confused and
00:00:53
simply stared at the formula and were
00:00:55
waiting for a magical aha moment where
00:00:59
the formula made sense I was always
00:01:01
asking myself how those smart scientists
00:01:04
could understand math so well they could
00:01:07
develop new algorithms and just think in
00:01:09
the language of math until I realized I
00:01:12
was thinking the wrong way around I was
00:01:15
focusing too much on the actual
00:01:16
mathematical formulas in the realm of
00:01:19
math instead of taking a step back and
00:01:22
thinking like a scientist I mean I was
00:01:24
literally looking at a formula and try
00:01:26
to understand the formula as a whole
00:01:29
which for me now now after learning this
00:01:31
secret just doesn't make any sense
00:01:33
anymore don't think of math like
00:01:35
something abstract and make it then
00:01:37
human interpretable you need to realize
00:01:40
that you need to think the other way
00:01:42
around think of the idea a human had
00:01:46
understand it and then think of how to
00:01:48
translate it into the language of math
00:01:51
this perhaps sounds very confusing but
00:01:53
math is not a standalone language in
00:01:55
which people think scientists think just
00:01:58
like you and me in natural language they
00:02:01
just know how to translate their ideas
00:02:03
into the formalisms of math which then
00:02:06
allows them to be implemented and then
00:02:08
further developed using the rules of
00:02:10
math as mentioned I was always looking
00:02:13
at a formula as a whole but each
00:02:15
component of a formula is just a
00:02:17
component of this human idea for example
00:02:20
a sum or a product are literally just a
00:02:22
for Loop that can have some conditions
00:02:25
that are literally equivalent to an if
00:02:27
else statement in code of course this is
00:02:30
easier said than done and to understand
00:02:32
a mathematical concept using human ideas
00:02:36
requires someone to actually properly
00:02:38
teach you these human ideas and how to
00:02:40
translate them step by step but in my
00:02:43
experience there are two scenarios one
00:02:45
the teacher does that already but you
00:02:48
don't understand he does it because you
00:02:50
were never explicitly taught to think
00:02:52
that way or two the teacher does really
00:02:54
only look at the formulas and
00:02:57
derivations in that case you need to try
00:02:59
to figure out the original human idea
00:03:01
Yourself by for example looking it up
00:03:04
online but the good thing is you now
00:03:07
know it is not your fault and there is
00:03:09
an intuitive understanding to the math
00:03:11
that you can find math is just the
00:03:14
formalization of a human idea very few
00:03:17
people actually think in the language of
00:03:19
math it's just a tool but when it comes
00:03:22
to every intermediate derivation step
00:03:25
then you often actually think in the
00:03:27
language of math which is very difficult
00:03:29
unless you know the next
00:03:33
secret this secret literally changed the
00:03:35
way I look at scary moth derivations
00:03:38
like this one again jumping back in time
00:03:40
to when I watched a lecture at College
00:03:43
my professor would explain the intuitive
00:03:45
idea of an ml algorithm show the
00:03:47
translation into the language of math
00:03:50
and then show us where we want the
00:03:52
formula to end up to make it more
00:03:54
efficient or simply actually work as an
00:03:57
algorithm that can be implemented but at
00:03:59
some point he would go off on a
00:04:01
derivation spree writing out one step
00:04:03
after the other and expected us to
00:04:05
understand why he did what he did
00:04:08
everyone was confused and of course
00:04:10
scared and annoyed but these derivations
00:04:14
are simpler than you might think not
00:04:16
easy but much simpler to execute after
00:04:19
learning this one secret I realized each
00:04:22
step was simply applying one specific
00:04:25
rule or a definition I realized that up
00:04:28
to a certain degree these mathematical
00:04:30
derivations or Transformations just
00:04:33
require you to have a list of rules and
00:04:35
tricks you need to collect that you can
00:04:38
then simply apply during the lectures
00:04:40
for each step I see I would explicitly
00:04:43
look for the rule and definition they
00:04:45
used and write that's down on my list
00:04:48
when solving or reading math derivations
00:04:50
on my own for homework assignments or
00:04:52
during an exam in most cases I would
00:04:55
literally just do some sort of pattern
00:04:57
matching I would look at where I current
00:04:59
L am go down my list of rules and
00:05:02
definitions and apply what fits the
00:05:04
pattern of course some patterns rules
00:05:07
and definitions are harder to spot than
00:05:09
others but after doing this for long
00:05:11
enough you just start to memorize
00:05:13
certain patterns but in general for most
00:05:16
ml math this secret technique does work
00:05:19
wonders you need to collect your
00:05:21
mathematical toolkit and learn to
00:05:23
recognize when you can apply which rule
00:05:26
which means practice practice practice
00:05:29
but math is of course far from all there
00:05:31
is to ml coding is also a very
00:05:34
challenging
00:05:36
skill learning the basics of python and
00:05:39
then an ml Library like pyo is really
00:05:42
cool and fun you simply follow a
00:05:44
tutorial online and have a really steep
00:05:47
learning curve you follow a recipe of
00:05:49
steps and Implement a lot of code you
00:05:52
really see and feel the progress you are
00:05:54
making but then when you want to go
00:05:56
further and learn to implement actual
00:05:58
algorithms or ml pipelines on your own
00:06:01
you hit a wall all of a sudden you sit
00:06:04
on one annoying problem for several
00:06:06
hours have written perhaps five lines of
00:06:08
code and you think you are not making
00:06:11
any progress this is very very
00:06:14
frustrating and is the point where a lot
00:06:16
of people determine coding is really
00:06:19
hard and that they will never be able to
00:06:21
really learn to code writing five lines
00:06:24
of code in 3 hours how pathetic I mean I
00:06:27
always thought I was so stupid for
00:06:28
writing code that never worked until I
00:06:31
debugged it for hours this can get so
00:06:34
bad that you don't even want to start
00:06:36
coding because you know it will fail but
00:06:39
that is simply the wrong way to think
00:06:41
actually writing code for 1 hour will
00:06:43
very likely mean let's say 3 hours of
00:06:46
debugging once I learned that this is
00:06:49
what it really means to be coding I all
00:06:51
of a sudden felt so relieved and just
00:06:54
not stupid anymore coding ml models
00:06:57
didn't feel impossible and hard anymore
00:07:00
because I was doing exactly what was
00:07:02
normal and expected and nowadays there
00:07:04
are amazing tools that I can't live it
00:07:07
out like GitHub co-pilot that can
00:07:09
generate code for you and explain code
00:07:11
for you but there's so much that you
00:07:13
simply learn through your own experience
00:07:16
or the experience of others that's why I
00:07:18
have a completely free Weekly Newsletter
00:07:20
where I share my experience as a machine
00:07:22
learning researcher including actionable
00:07:24
tips AI news and more I'll just pin a
00:07:27
comment below with the link to sign up
00:07:30
but anyway you have to realize that
00:07:33
writing code is not actually coding
00:07:35
debugging is coding this realization
00:07:38
really helps you with implementing
00:07:40
things on your own step by step but when
00:07:42
you have to work with an existing code
00:07:44
base where you have everything at once
00:07:47
you will probably still be overwhelmed
00:07:49
so let's look at the next secret
00:07:53
tip there are two cases where you will
00:07:56
need to understand complex code the
00:07:58
first one is when build bu on top of an
00:08:00
existing repository I remember when I
00:08:02
started working on my first larger
00:08:04
project where I built on top of an
00:08:06
existing code base it was so much code I
00:08:09
literally had no idea where to start
00:08:11
just like in the previous secret I again
00:08:13
felt like learning to read code was as
00:08:16
impossible as writing code all those
00:08:19
tutorials and smaller personal projects
00:08:21
didn't prepare me for this much code I
00:08:23
Tried reading each source file and
00:08:25
started to write code as soon as
00:08:27
possible into places I thought made
00:08:29
sense but that unsurprisingly LEDs to a
00:08:32
lot of headaches and wasting time
00:08:34
writing code that was destined to fail I
00:08:36
had to learn the hard way that there is
00:08:39
a very simple strategy to understanding
00:08:42
large code bases I wish someone would
00:08:44
have simply told me once how to approach
00:08:46
a challenge like this with most large ml
00:08:49
repositories you have a train.py and an
00:08:52
evil. piy file those should always be
00:08:54
the starting point I find those files
00:08:56
set a breakpoint in the beginning and
00:08:58
start stepping through the code with the
00:09:02
debugger I cannot emphasize enough how
00:09:05
simply yet insanely effective this
00:09:07
technique is it's literally like
00:09:09
cheating you can step through the data
00:09:11
pre-processing training Loop the actual
00:09:13
model the evaluation metrics and every
00:09:17
other detail depending on the code base
00:09:19
and your experience this takes just a
00:09:22
few hours and you have an amazing
00:09:24
overview of the code base and will have
00:09:26
a much better feeling for where to add
00:09:29
the new code for your own idea that said
00:09:32
you might not always want to build on
00:09:34
top of an existing highly optimized code
00:09:36
base but simply want to understand an
00:09:39
algorithm better for example when you
00:09:40
want to understand poo a famous
00:09:43
reinforcement learning algorithm I would
00:09:45
not recommend to look at the optimized
00:09:47
implementation that's way too Overkill
00:09:49
and complex luckily for many important
00:09:52
models there are minimal educational
00:09:54
implementations that just implement the
00:09:56
main idea so that people can understand
00:09:59
the model and here yet again the best
00:10:01
way is to set a breakpoint in the
00:10:03
beginning of the main function and then
00:10:06
just start debugging finally there's one
00:10:08
fundamental secret to mastering machine
00:10:10
learning that you need to
00:10:15
know this final secret ties everything
00:10:18
we just discussed together and is the
00:10:20
one reason that will determine your
00:10:22
success or failure of mastering ml 34%
00:10:26
of organizations consider poor AI skills
00:10:29
expertise or knowledge as the top reason
00:10:32
blocking successful AI adoption
00:10:34
according to an IBM study from 2022 why
00:10:37
do you think people fail learning ML and
00:10:40
fall into the category of people with
00:10:42
poor AI skills is it because it is hard
00:10:46
yes but it was also hard for every
00:10:48
person who has now mastered it people
00:10:50
fail to master ml because they stop
00:10:53
learning machine learning too early and
00:10:55
give up and why do people give up
00:10:58
because they have false expectations and
00:11:01
don't enjoy the process of learning they
00:11:03
think mastering ml is hard because they
00:11:06
didn't learn ml in a few weeks or
00:11:08
because they didn't understand a video
00:11:10
explaining an ml concept the first time
00:11:13
they will never understand it I took my
00:11:15
first introductory AI College course
00:11:18
about 3 and a half years ago after that
00:11:20
semester I took my first real ml course
00:11:23
along with working on my first ml
00:11:25
projects after that semester I continued
00:11:28
with my first deep learning courses and
00:11:30
continued working on projects and
00:11:32
reading a lot of papers I definitely
00:11:35
didn't understand everything the first
00:11:37
time but I knew that was normal over
00:11:39
time I learned the secrets mentioned
00:11:41
before and that somewhat mastering ml
00:11:44
takes time I failed my first ml
00:11:46
interviews for internships at Amazon
00:11:48
nuro and Google deep mind but now I am
00:11:51
working with an ex meta professor and
00:11:54
collaborating with a Google deepmind
00:11:56
researcher it takes time period this is
00:12:00
not a skill you learn over a weekend the
00:12:02
10,000 hour rule applies here as well if
00:12:05
you spend 10,000 hours on a specific
00:12:07
skill you will master I'm absolutely not
00:12:10
trying to discourage you rather the
00:12:12
opposite I don't want to be some weird
00:12:14
influencer that wants to sell you a
00:12:16
dream of mastering ml in a few weeks I
00:12:19
want to encourage you to really learn
00:12:21
machine learning to really learn the
00:12:22
theory and really gather practical
00:12:25
experience and the Mastery of machine
00:12:28
learning comes after after learning the
00:12:29
Fundamentals by really working on
00:12:31
projects encountering real world
00:12:33
problems and reading real
00:12:35
state-of-the-art papers or respective
00:12:38
blog posts by having this expectation
00:12:40
that it will take time you relax way
00:12:43
more and the learning process becomes
00:12:45
much easier enjoyable and successful in
00:12:48
the end all these secrets are
00:12:50
universally true no matter how you
00:12:52
decide to learn machine Lear and I say
00:12:54
that because there are mainly three ways
00:12:55
to do so and if you want to find out
00:12:58
which one is best for you you might want
00:13:00
to watch this video next I there talk
00:13:02
about data science but everything
00:13:04
applies the same way to machine learning
00:13:06
in
00:13:07
general
00:13:09
bye-bye
