6 Years of Studying Machine Learning in 26 Minutes
Summary
TLDRThis video script narrates the journey of a machine learning enthusiast, detailing their six-year transformation from a computer engineering student to a research scientist at a favorite AI startup. The speaker shares their academic and professional experiences, including initial struggles, pivotal courses, research projects, and the process of landing their dream job. The script also highlights the importance of continuous learning, making informed career choices, and avoiding common beginner mistakes in the field of machine learning.
Takeaways
- 😀 The speaker's journey in machine learning started with an interest in physics and math, leading to a Computer Engineering degree at TU Berlin.
- 📚 The initial years of university were challenging, focusing on foundational courses like linear algebra, calculus, and differential equations, which later became crucial for understanding machine learning.
- 🔧 The speaker's first job as a student researcher in an optical physics lab helped build a strong foundation in experimental work and programming, but also highlighted the importance of seeking new challenges to avoid stagnation.
- 💡 The discovery of AI courses in the fifth semester marked the beginning of the speaker's deep dive into machine learning, starting with reinforcement learning.
- 🎓 The bachelor thesis on deep reinforcement learning for autonomous robotic navigation was a significant step, despite the initial struggles and the steep learning curve.
- 📈 The transition to a Computer Science master's program allowed the speaker to focus exclusively on machine learning, taking advanced courses and engaging in more complex projects.
- 🤖 Working on robotics projects and publishing a paper at a top conference like IROS was a major milestone, showcasing the speaker's growth and capabilities in the field.
- 👨🏫 The speaker's experience of switching teams within the research institute to work in the AI department highlights the importance of proactively seeking opportunities to learn and grow.
- 📘 Reading papers and learning from them independently became a habit that contributed to the speaker's development as a researcher.
- 🚀 The pursuit of internships and jobs at top companies, including rejections and learning from them, demonstrated resilience and a commitment to continuous improvement.
- 🌟 The speaker's eventual success in joining a favorite AI startup as a research scientist after years of hard work and learning emphasizes the value of persistence and incremental progress.
Q & A
What was the speaker's initial academic background before getting into machine learning?
-The speaker initially studied Computer Engineering at TU Berlin, taking courses in linear algebra, calculus, differential equations, and basic programming in C.
How did the speaker's interest in machine learning begin?
-The speaker's interest in machine learning began during their fifth semester when they chose their first AI course, which was split into two parts: old school AI and reinforcement learning.
What was the speaker's first experience with a machine learning project?
-The speaker's first machine learning project was their bachelor thesis, where they worked on deep reinforcement learning for autonomous robotic navigation.
What was the speaker's first job as a student researcher, and how did it relate to their later work in AI?
-The speaker's first job as a student researcher was at an optical physics lab, running experiments with optical fibers. Although not directly related to AI, it provided them with basic programming skills that later became useful in machine learning.
How did the speaker's experience with coding evolve throughout their studies?
-The speaker started coding in C during their computer engineering studies, then moved to Java for data structures and algorithms, and later used Python for machine learning projects, including using libraries like PyTorch.
What is Codium, and how did the speaker use it to improve their coding efficiency?
-Codium is a free coding assistance tool similar to GitHub Copilot. The speaker used it to refactor and explain existing code and to write appropriate functions using context from their entire project.
What was the speaker's approach to learning new machine learning concepts?
-The speaker learned new machine learning concepts through a combination of university courses, self-study using YouTube videos, reading papers, and hands-on projects.
How did the speaker's job change when they switched teams within the research institute?
-The speaker transitioned from working in an optical physics lab to joining the AI department, where they started working on data engineering and gained more experience with machine learning.
What was the speaker's experience with internship applications, and what did they learn from it?
-The speaker faced rejections from several top ML internships but learned the importance of persistence and self-improvement. They eventually secured an offer from their favorite AI startup.
What is the speaker's advice for beginner ML students to avoid common mistakes?
-The speaker suggests that beginner ML students should watch a follow-up video where they share seven common mistakes to avoid, emphasizing continuous learning and improvement.
What was the speaker's final decision regarding their academic and career path after completing their master's degree?
-The speaker decided to join their favorite AI startup as a research scientist instead of pursuing a PhD, although they had an offer for a PhD position.
Outlines
🎓 Journey to Becoming an AI Research Scientist
The speaker reflects on their six-year journey in machine learning, starting as a Computer Engineering student at TU Berlin with no knowledge of AI. They describe the initial struggle with foundational courses like linear algebra and calculus, which later became crucial for understanding machine learning models. The speaker also shares their first steps into the tech industry through a student researcher role at an optical physics lab, emphasizing the importance of continuous learning and the steep learning curve of new jobs.
📚 Transitioning from Academics to Practical Machine Learning
The speaker discusses their transition into machine learning during their third year at university, where they took their first AI course, which included reinforcement learning. They recount the challenges of their bachelor thesis on deep reinforcement learning for robotics, which involved self-teaching and overcoming hardware setup difficulties. The paragraph also highlights the speaker's first publication at a top robotics conference, IROS, marking a significant milestone in their academic career.
🔬 Balancing Theory and Practice in Machine Learning
In this paragraph, the speaker delves into their fourth year, focusing on machine learning courses and projects. They mention taking a classical machine learning course, learning about various algorithms, and gaining practical experience through coding assignments. The speaker also talks about their experience working at a research institute's AI department, where they learned data engineering and gained more experience with PyTorch. They emphasize the steep learning curve and the excitement of working on new ML domains like computer vision for medical image analysis.
🚀 Pursuing Advanced Studies and Research in Machine Learning
The speaker recounts their graduate studies, where they took advanced courses in deep learning and computer vision, and worked on various projects, including one that was published but never successfully. They express a desire for a more dynamic learning environment, leading to a job change to work on graph neural networks. Despite the initial rejection, they eventually joined the ML department, where they continued to learn and grow in the field of AI.
🛠️ Overcoming Challenges and Exploring Multimodal Learning
The speaker describes their quest for a more challenging and dynamic work environment, leading to applications for internships at top tech companies and startups. They recount their experiences with interviews and the process of finding a professor for their Master's thesis in multimodal learning. The paragraph highlights the importance of persistence and the pursuit of research interests, even when facing rejections and setbacks.
🌟 Securing a Position at a Dream AI Startup
In the final paragraph, the speaker shares their decision to apply for research scientist positions despite the risk of rejection. They detail the application process for a favorite AI startup and their surprise at receiving an invitation for an interview. The speaker reflects on their growth and the importance of daily improvement, ending with the announcement of their upcoming role at the AI startup and an invitation to watch a follow-up video on common mistakes for beginner ML students.
Mindmap
Keywords
💡Machine Learning
💡Research Scientist
💡Computer Engineering
💡Mathematics
💡Coding
💡Reinforcement Learning
💡Deep Learning
💡Convolutional Neural Networks
💡Multimodal Learning
💡Technical Papers
💡Internship
Highlights
Six-year journey in machine learning, starting from an ML student researcher to joining a favorite AI startup as a research scientist.
Initial introduction to technology through a combination of interest in physics, math, and the practicality of engineering.
Foundation in computer engineering at TU Berlin, with early struggles in understanding the relevance of mathematical concepts.
The importance of math skills as a fundamental for later machine learning understanding and intuition.
First job as a student researcher in an optical physics lab, highlighting the initial steep learning curve and eventual plateau.
The transition from computer engineering to focusing on machine learning during the third year of university.
First exposure to AI through a course on reinforcement learning, which sparked interest in machine learning.
Bachelor thesis on deep reinforcement learning for autonomous robotic navigation, marking a significant step into deep learning.
The experience of publishing the first ML paper at a top robotics conference, IROS.
Switching to a pure computer science master's program to delve deeper into ML, facing challenges with engineering and training ML models.
The process of learning to read and understand ML papers, a skill that became essential for research.
Internal job switch within a research institute from physics lab to an AI department, highlighting the importance of persistence.
Exploration of computer vision and medical image analysis, introducing the challenges of learning new domains.
The realization of the need for continuous learning and improvement in ML, as well as the decision to pursue a PhD.
The pursuit of internships and the challenges faced in the competitive tech industry, including rejections and learning from failures.
Finding and working with a professor specializing in multimodal learning, emphasizing the importance of finding the right mentor.
Research on video moment retrieval in the field of multimodal learning, achieving state-of-the-art performance.
The decision to join a favorite AI startup as a research scientist, illustrating the culmination of years of hard work and dedication.
Reflection on the importance of daily improvement, enjoying the work, and being proud of achievements in the ML field.
Transcripts
I've been studying machine learning for
the past 6 years in which I worked as an
ml student researcher for over 2 years
and have written my first three papers
my journey started studying Computer
Engineering not knowing what ml was or
even that it existed to where I am now
soon joining my favorite AI startup as a
research scientist in this video I'll be
sharing my experience over these six
years and explain what I did each year
to get where I am now things like what
to expect for the first few years what I
did to get my first ml student roles and
most importantly what you should be
avoiding and trust me there's a
lot okay before we get to years 1 and
two how did I get into Tech well young
Boris liked physics and math in high
school and thought hm with physics you
can't really make money so I need to do
engineering
I.E applied physics building a robot
would be really cool but then I also
need to know how to program the robot to
make it do the stuff I wanted to do at
that time I didn't know Ai and ml
existed but those were my thoughts those
led me to studying Computer Engineering
at the TU Berlin the first two years
were really tough of course I had to
take the standard courses like linear
algebra 1 Calculus 1 and two and a
course on differential
equations luckily I genuinely enjoy
learning math but that doesn't mean it
was easy for me in the beginning it all
doesn't make much sense and you don't
know why you are learning all these
mathematical formulas and Abstract
Concepts but I promise you at some point
most of them will make sense and you
will learn to appreciate and make use of
them especially when learning ml the
basics of these math skills will be the
fundamentals you will later need for ML
and give you an intuition for how to
look at ml models in a mathematical
sense but back then again I didn't even
know AI existed I had a lot of
electrical engineering and even some
Physics courses those were more tough
but I also had my first computer science
courses and learned how to code in C yes
in C that's right remember I was a
computer engineering major so my program
was designed for low-level coding and
electrical
engineering but I still had the standard
courses on data structures and
algorithms in Java and also a course on
theoretical computer science all that is
pretty much the standard things you
learn when getting into computer science
related programs some CS Theory and a
lot of coding luckily coding is much
easier nowadays with a cool coding
assistance out there a tool I cannot
live without anymore you have probably
heard about GitHub co-pilot but what if
you want to use a tool that has pretty
much the same features but you could use
it for free as an individual well that's
where today's sponsor comes into play
codium you can very easily install it in
any of your favorite IDs and then simply
hit command plus I and type in what you
wanted to do the cool thing is it can
not only refactor and explain existing
code but it also uses a lot of context
from your entire project to figure out
how to write the most appropriate
functions using different apis spread
out throughout your code base it looks
at your open tabs and even the entire
git Repro you can even go beyond that
and add links to other files and Repros
to use those as context as well so to
improve your coding efficiency use
codium for free forever using my link in
the description below no strings
attached just an amazing free product
but now let's get back to what I did
besides normal college courses Landing
my first student researcher job at an
optical physics lab about 6 months into
my first year I wanted to somehow boost
my resume earn some money to survive
college and also just learn more stuff I
then saw this listing at a research
institute directly next to my uni and
applied it was honestly quite surprising
that they invited me to an interview
because I literally had not much to
offer except basic programming skills
but I guess for the job I was supposed
to do it was enough I was responsible
for running a lot of experiments with
Optical fibers and doing measurements
when starting a new job or project the
learning curve will likely be very steep
which is amazing I learned a lot but if
you do the same measurements for over
one and a half years the learning curve
plateaus and the job becomes boring in
total I stayed at this job for 3 years
and this learning curve was completely
flat after perhaps 8 to 9 months if not
less and this was a big mistake I really
should have at least changed to a
different team at this Research
Institute after a year but I was quite
exhausted for these first two years I
slept 6 hours at night didn't do much
Sports and just worked a lot which is
normal I don't want to PL I had a lot of
fun in fact I am happy and proud I did
all that but yeah all of this happened
in my first two years of uni most
importantly I learned the basics of math
and computer science and worked as a
student researcher all of which helped
me with studying machine learning
without even knowing ml
existed I finally got into the third
year semesters five and six where I
could choose some of my courses myself
this fifth semester is where I saw that
AI courses existed at my uni and is
where I chose my very first AI course
this is where the ml Journey really
started that said this AI course was
split into two parts the first one was
about old school AI not machine learning
yes AI does not necessarily mean ml if
you have an algorithm with a set of
rules to make decisions it's effectively
AI I learned about things like the
strips method looking back it's not that
exciting honestly but that is where I
started and back then I thought it was
decently cool but the second half of
this course was really cool the second
half was about reinforcement learning
which in retrospect is a weird start
into ml learning about RL before even
knowing what a newal network was but
maybe this is a good way to show you
that it does not really matter how you
start if you keep going you will learn
all the fundamentals anyway just in a
different order perhaps but I would
still not recommend it if you have the
option to choose but you know you you
get the point anyway I learned about
things like Bandit Theory Monte Carlo
research marof decision processes and
finally RL algorithms like Q learning so
in my fifth semester 2 and a half years
into college there was still not that
much ml but these RL lectures really got
me interested in ml and especially RL
that's why I wanted to do my bachelor
thesis in RL which is what I did in my
sixth semester I worked on deep
reinforcement learning for autonomous
robotic
navigation this was a complete cold
start into deep learning I didn't even
know what a new network was I had to
learn all of that on my own through
YouTube videos even worse in the
beginning I struggled a lot to even get
the hardware set up and when I reached
out to my supervisor for help he said he
thought I might not be ready for this
thesis and I had two weeks to prove him
otherwise and if I failed he would have
to drop the thesis with me which would
have been so bad the semester had
already started and then I would have to
look for a completely new one but I
pushed through
and made it this thesis project was a
lot of work a lot of engineering work
and no real training itself since the
thesis was more on the deployment side
of DRL agents than the training side
nevertheless I still learned a lot of
core coding skills like debugging and
did get to learn pyo for the first time
so my final Bachelor year was still a
slow step into the world of ml but a
very firm one one that set the path to
going all in on ML which is why I then
switched to a pure computer science
master so my fourth year began and I
went all in on ML I selected only ml
courses and projects but this of course
came with a lot of challenges in my
first graduate semester I pretty much
had one big course and one big project
for the project I continued to work on
the same team for autonomous robotic
navigation that I worked with during my
bachelor thesis the project was still
more of an engineering effort because we
built a benchmarking suite for
autonomous robots which again came with
a lot of failing and debugging but this
time I could focus a lot more on
training our own agents using pytorch
and had to start reading papers to learn
things like Po of course the beginning
of reading papers is always a bit tough
because you have to get used to the
lingo but I felt so cool I felt like a
real scientist the really cool thing was
that later that year we actually
published this work to one of the two
best robotics conferences iros that was
so huge for me it was my first ml paper
and it was even published at a top
conference now alongside this project I
had my first real ml course I learned
all the basics of classical machine
learning for example what is supervised
learning unsupervised learning what is
the bias variance trade-off what are
methods like linear regression decision
trees support Vector machines K means
PCA boosting and Ensemble learning and I
learned about all the basics of new
networks like what loss functions grade
and descent back propagation and
regularization are alongside each
lecture there were of course practical
homework assignments to implement the
ideas we learned during the lecture and
those were again using py do now besides
uni I still had this boring physics lab
job at this point I was working there
for two to two and a half years already
but the cool thing was the research
institute I worked at also had an AI
department so I wanted to internally
switch teams I applied got an interview
and was rejected I mean I get it I was
just starting my first real ml course
and had no theoretical knowledge of any
of the ml fundamentals so I tried again
half a year later after completing the
ml course and having gathered more basic
pyto experience and then actually did
get the job what an amazing start to my
second graduate semester my second half
of my fourth year I started my work as
an applied scientist student researcher
in the ml Department I again had a steep
learning curve and was so excited to get
to work these first six months I started
working on a lot of data engineering
mainly using pandas which I have never
used before I learned a lot there and at
Uni I also focused on Purely practical
learning I took two project courses I
again continueed to work on this
robotics project but at this point I
felt a bit more of a fatigue working on
the project it wasn't that exciting
anymore but it still was a lot of work
and my learning curve PL toed but I
contined to work on it because I hoped
for another paper nevertheless I started
to look at other cool ml domains and
took another project course a project on
computer vision for medical image
analysis this was my first computer
vision project and I had to detect
anisms in 3D images of the brain it was
really cool but I have never dealt with
computer vision before and had never
learned what a convolutional new network
was so the learning curve was again very
steep I had to learn all of that
knowledge Myself by watching YouTube
videos and reading more papers in the
end the final project was not the worst
but also not the best either at least
looking back at it now and I think this
is a good thing if you are looking back
at Old projects and think they are bad
and and perhaps even cringing because
you would have done things differently
with your current knowledge then you
have gotten
better so yeah this year was packed with
all the ml I could fit in most of it was
actually working on ML projects and only
taking one ml lecture but a really
important one so far it was quite
straightforward but in my next year I
had to make some important decisions
now uni continued as usual but
career-wise I had to make those
important decisions in my third graduate
semester I again took one lecture course
and two more projects I took my first
actual deep learning course which had a
decent overlap with my first ml course I
again learned about the same
fundamentals of new networks but now
also had lectures on cnns recurrent new
networks Auto encoders and a bit of
explainable AI so nothing too crazy
right at this point I am really into AI
myself and I started watching paper
review videos on YouTube and reading
random papers on my own perhaps because
this course didn't have too much new
stuff and my job didn't teach me much
theoretical content as well but anyway
this habit of reading papers and
learning stuff on my own are things I
still do to this day and that I I
genuinely enjoy so besides this deep
learning lecture I once again worked on
this robotics project and I have to say
working on it this semester just wasn't
necessary it was really not that
interesting anymore and I really just
wanted to learn new stuff but I was
still hoping for a paper which in the
end was never successfully published now
my second project course this semester
on the other hand was again about
reinforcement learning but was amazing I
had to thoroughly read a paper and
actually reimplement it and reproduce
its results which was a lot of fun I
often say it and I'll say it again
reimplementing a paper and recreating
its results is one of my favorite
projects to recommend I even wrote a
blog post about it and submitted it to a
top ml conferences blog post track but I
didn't really know how the process
worked back then and
I did get my reviews but never received
an email telling me that they were
released so when I randomly checked I
saw the reviews and that I never
responded to them thus the article was
rejected from the icr blog post track
nevertheless the project taught me a lot
and at this point I was pretty confident
I wanted to become a top ml researcher
and this goal meant for me I needed to
strive for the best companies my job at
that time as a student researcher was
not completely plateauing but also not
the best anymore we started doing
research on graph newal networks but for
over a year now we were still stuck with
a lot of the same boring data
engineering and feature engineering I
effectively didn't really learn anything
new that's why I wanted to find a new
job and not make the same mistake as
before where I stayed for three years at
the same job so I applied to dozens of
top ml internships and I actually got
invited to an interview for a applied
science internship at Amazon that was my
first real Tech interview it was really
exciting except that I failed miserably
the more frustrating part was that the
questions were really not that hard it
was a rapid fire basic ml questions
interview they were literally asking
about the content of my first ml course
I mentioned before the one I completed
not even a year ago but well life goes
on and I got another interview at a cool
startup called nuro this time it was for
an ml engineering internship and the
first interview round was a coding
interview again something completely new
to me I prepared using lead code but
when I saw a blank code in canvas and no
pre-existing code where just just had to
fill in an algorithm I was so scared I
failed miserably again well the
applications weren't going so well I
simply didn't get many more interviews
so I changed my Approach I directly
reached out to a Google Deep Mind
researcher I found interesting and asked
for an internship and he got back to me
we had an interview call where I felt it
went decently well but I got rejected I
was done with looking for internships
and focused on finding a new job as a
student researcher where I could also do
my Master's thesis I decided I had
enough of reinforcement learning and
found computer vision really cool but
then I thought how cool would it be if
you could talk to an AI about images or
even
videos well that's where I decided
multimodal learning was really cool but
at my University there was a problem
there were no professors working on
multimode learning and pretty much all
of the professors were how do I say it a
bit more old school and not that much
into the new stuff there definitely were
one or two don't get me wrong but they
just weren't into something really
similar to multimod learning so I looked
outside of my uni tiim I wanted to look
for a professor that was a bit more
active and ambitious I read multimode
learning papers and looked at the
authors I then Googled them to see if
they could be an option as an adviser
for my research and thesis and then I
found the perfect Professor he was young
was just about to start as a professor
and before that he was a postor at UC
Berkeley and a researcher at meta and he
worked on multimodal learning he was
everything I was looking for long story
short I am so happy to have gotten the
job and started to work with him later
in my final year I still had my goal of
getting to Big Tech but there are these
nice sayings Rome wasn't built in a day
and all roads lead to Rome I.E
everything takes time and there are
multiple ways to get where you want to
get so all in all this semester besides
this career hassle I just did a lot of
coding at my job for the robotics
project and for this RL paper
reimplementation but this was still just
the first half of my fifth year my
second half was not that eventful since
I failed all my applications for summer
internships I was still doing my best to
learn stuff at my at the time current
job otherwise not much interesting stuff
happening there and at Uni I really
focused on computer vision I took a
course on automatic image analysis and
another seminar course on deep learning
for computer vision where I had to read
several papers on self-supervised
learning and presented them to the group
that was so much fun I just really enjoy
reading papers I even made my
presentation into a mini YouTube series
on self-supervised learning but besides
those two courses I took my second
General deep learning course this one
was finally a bit more advanced I
learned about things like representation
learning self-supervised learning
Transformers Gans diffusion models graph
new networks and even neural ordinary
differential
equations and finally I also did another
computer vision project course where I
wrote a paper/ techical report on so
there was way more theoretical content
this semester but still a practical
project now you might have noticed that
this semester usually should have been
my final semester usually the Masters
would end after 2 years but I had
actively decided to give myself one more
year mainly to have one semester for an
internship and one more for my thesis so
this semester was my last one with
courses and since I didn't get an
internship I had one more entire year to
focus on doing research with my new
professor and then completing my
Master's
thesis and that is what I did in my
final year
I was finally done with uni at least it
felt like that because I had no more
exams I started working as a student
researcher with this cool professor and
started doing research on multimodal
learning specifically video moment
retrieval I read a lot of papers
developed a model that achieved new
state-of-the-art performance on the
benchmarks I evaluated on and wrote a
paper on it in a very short time I even
submitted the paper to a top conference
and I'm telling you those were some
stressful weeks but it still recently
got rejected and to be honest I probably
understand why I rushed it because we
chose a deadline that was simply way too
close I should have taken more time and
just submitted it to a later conference
so that the paper was overall more solid
although it was annoying I will continue
to improve this work and soon submitted
to another conference then I remembered
that I'm still in my final year I still
need to actually complete my degree LOL
that's why I'm currently still in the
process of finishing to write my thesis
and handing it in but since this is my
final year I also had to think of what
comes next I thought to myself either I
skip the PHD and become a researcher at
a top lab or I do my PhD D I mean How
likely was it to skip the PHD the cool
thing was I already had an offer from my
professor for the PHD position and I was
very happy to accept it nevertheless I
still want to try out applying to two
companies as a research scientist one
was deep mind and although I thought my
chances were in fact decent because I
had exactly the combination of different
experiences that they were looking for I
got rejected but besides deep mind I
applied for another really cool AI
startup my favorite one to be precise I
knew I wouldn't even get invited to an
interview but one evening I was like why
not they won't invite me
anyway but you probably already know
where I'm going with this they did
invite me and I was shocked the
application process was quite tough and
I wanted to really give it my all and
see if I am good enough for them and
well long story short I did get an offer
and will work for them starting in a few
months at the time of this recording
once I start my work I will announce
which company it is don't worry I just
want to make it cool because for me it
is a big thing but yeah anyway
throughout all these years there was a
lot of struggling but also some
occasional successes I quickly learned
that the important thing is to keep
moving think some people get to where I
am now in less time and some in more but
that doesn't matter what matters is that
you try to improve every day by 1%
overall enjoy what you do and that you
are proud of what you do nevertheless
there are many mistakes you can avoid
and not waste any time on if you simply
know what they are that's why you might
want to watch this video next I there
share seven common mistakes beginner ml
students make every year happy learning
and bye-bye
Ver Más Videos Relacionados
How I Became A Data Scientist (No CS Degree, No Bootcamp)
TEKNIK INFORMATIKA
How to Start Coding in 2024? Learn Programming for Beginners | Placements & Internships
Advice From a Top 1% Machine Learning Engineer
Machine Learning Projects You NEVER Knew Existed
Come sono diventato MACHINE LEARNING ENGINEER in GLOVO | Guida Step By Step
5.0 / 5 (0 votes)