8 Mistakes Holding Your Career Back | Machine Learning

I reached out to some of the most prominent voices in the Machine Learning community:
- François Chollet
- Soumith Chintala
- Laurence Moroney
- Bojan Tunguz
- Sebastian Raschka
I asked them a simple question:
What are the most common mistakes people make when starting in machine learning?
This video is their answer.
🔔 Subscribe for more stories: www.youtube.com/@underfitted?...
📚 My 3 favorite Machine Learning books:
• Deep Learning With Python, Second Edition - amzn.to/3xA3bVI
• Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow - amzn.to/3BOX3LP
• Machine Learning with PyTorch and Scikit-Learn - amzn.to/3f7dAC8
Twitter: / svpino
Disclaimer: Some of the links included in this description are affiliate links where I'll earn a small commission if you purchase something. There's no cost to you.

[@hiankun]

"Better data is better than better models!"

[@underfitted]

Yessir!

[@yinli1240]

That’s why I don’t trust the data analysis linked to Covid and climate changes

[@RafaRodriguezOfficial]

Reminds me of something Andrew Ng teaches in his ML class, the first question he asks when they hire him to improve a model is how much would it cost to have 10x more training data.

[@douglasdbs7139]

I started studying machine learning in 2009, so before the boom of deep learning in 2012 after the ImageNet competition (at that time, the term "statistical learning" was still being used around). I remember that since that time my PhD advisor used to say that "nowadays people are too focused on overtuning SVM or HHM-DNN models, but at the end these models are extrapolating hypotheses from data, so data should be the most important thing you should care in your problem". I have been preaching for a more data orientend ML/AI for years. To be honest, in 2015, 2016 it looked like a lost cause. I was super excited a few years ago when I learned that Andrew Ng started this data-centric AI moviment. I could not agree more that "better data is better than better model". Also, I agree that very often you can build a very decent ML pipeline without using deep learning at all, just by careful crafting your data and knowing how to measure performance and impact. Learning information Theory and Stochastic Processes to this end in particular is super helpful (knowing how to interpret and measure uncertainty).

[@arijitgoswami3652]

100%

[@pavlobilinskyi775]

When I started to learn machine learning three years ago, I had a good understanding of theory, but in practice I was helpless like puppy. Even simple toy Kaggle competitions made me struggle and spend days of my time. I started to think that ML is extremely hard and I will never succeed at this. But as I started working as Data Analyst, I learned to do problem-oriented approach to the task, started searching for simple solutions. Thank you for sharing these important points! I wish someone had told me back then.

[@underfitted]

Awesome story, man! You are absolutely right. Practice is crucial.

[@MohamedSalah-dr2qd]

Wish you Good luck pavlo, do you think, data analytics is a good gateway to master machine learning or data science later?

[@pavlobilinskyi775]

@@MohamedSalah-dr2qd Hi, Mohamed! It may be not the shortest way, but it really gives you opportunity to become Data Scientist. I'm from Ukraine, in my country it's difficult to find Data Science job for entry-level guy, but there are more jobs for entry Data Analyst. If you don't have experience working with real-world data, working as Data Analyst will help you. Basic analytical skills, methodology of working with data, SQL, AB-tests will give you a solid basis to grow to Data Scientist. However, there is still a huge gap between these two roles, so you will need to learn and practice on your non-working time. It would be great to find some opportunity (classes on ML, some internship, DS events), where you can meet Data Scientist's and get advice from them. Wish you good luck!

[@ponugotimanojkumar]

How did you learn machine learning from basics?any tips for me to learn ML as a beginner especially with a medicine background (as I'm a doctor) Thank you

[@abhishekgrewal8906]

Proud to have taken Intro to ML under Sebastian Rascha during my undegrad at UW Madison. Must say he is a great professor, and a lot of things he taught during class, I find useful working on projects (though I am guilty of dosing out during the lectures lol).

[@sehrishilyas8416]

Thanks for the motivation to not get overwhelmed by these mistakes and to get rid of them!

[@underfitted]

Absolutely!

[@raul825able]

Wow, I follow your twitter regularly but somehow missed this video. I stumbled upon it just today and man, it's a gem!! Thank you!

[@mustacheman6000]

The point about better data is super interesting, do you have a video talking about where to get better datasets or some sort of resources outlining strategies to find it?

[@JasonAndrewsUK]

Another brilliant video. Thanks. I found the audio of the percussion instrument a little distracting but once again, superb video, well executed. Your channel will explore with videos like this and it's well deserved. Making me want to get into ML!

[@underfitted]

Thanks! Yeah, I’m working hard to improve with the audio. The last couple of videos are much better. Thanks Jason!

[@1littlecoder]

That's great editing. Love the way this video is made!

[@underfitted]

Glad you liked it! Really appreciate your comment!

[@MrBratkenSolov]

ah yes, choppy jittery videos are pinnacle of editing

[@bhapapathvisutvatanasak9329]

Very educated! Thank you very much for collecting these advices!

[@underfitted]

Your welcome!

[@shoaibsh2872]

Came here from your Twitter. Didn't know you have such quality KZbin channel, instant sub 🤝

[@underfitted]

Thanks so much!

[@zulfiquarali710]

I so badly needed this video. God bless you 🙏

[@inventorokoth]

Nice work, thanks for the insights. I can say that the point of going full throttle is indeed true because for me i have been learning React and did not go through hours of tutorial but just set out to build one complete web application and got right into it without any prior React knowledge. I can testify I have learnt a tonne of stuff and even as i get into ML and AI, i'll surely employ what I have learnt from this video

[@jespermikkelsen7553]

Love your videos - and your style, humble about the subject...🙏

[@underfitted]

Thanks Jesper!

[@Darkev77]

Awesome video! Can you elaborate on what you exactly meant by “A metric without context means very little.”? What kind of metric-context relationship would make sense?

[@underfitted]

Thanks, Darkev! I built a model and it has 95% recall. Is it a good model? Well, it might be… but what about if I tell you that humans can easily achieve 100% recall on the same task? Then 95% is not that good anymore. That’s what I mean by context: a metric alone is not enough. Does that help?

[@Darkev77]

@@underfitted yes, awesome, thanks a lot!

[@satishkumar19871]

Fruitful information for ML/DL beginner

[@underfitted]

Glad it helps!

[@71sephiroth]

Thanks for the awesome video! Mind sharing the full chat with Chollet? He is one of those rare guys where the SNR ratio is 1.

[@underfitted]

Ping me on Twitter @svpino and I’ll share it

[@mgreek31]

WOW !!! I needed this . Thanks Santiago

[@underfitted]

Glad it helps!

[@sandeepbansal9555]

that's very informative! :) Great Video 👍

[@omarmahtab6851]

One of the best ML videos, thanks man!

[@underfitted]

Thanks Omar!

[@dexkconsultoria-solucoesem7296]

today i've learn a lot here, thank you.

[@davinjimeno5408]

your videos are really valuable for people new to ml

[@nimitkanani1691]

Very Very Helpful.

[@dheerajkrovvidi7199]

Thanks for the tips Mr. Santiago

[@underfitted]

Your welcome!

[@andyowning11]

Great videos. Keep it up!

[@underfitted]

Thanks, will do!

[@connorhawkins2363]

Amazing video. Well done!

[@underfitted]

Thanks!

[@gustavojuantorena]

Great video!

[@underfitted]

Glad you enjoyed it, Gustavo! Thanks!

[@The64v]

"Books are relevant, but they are ust one piece of a bigger puzzle. You gotta do the thing."

[@jeanded]

Great video :)

[@underfitted]

Thanks, man!

[@dataflex4440]

amazing content

[@GEMSofGOD_com]

It's crazy how enthusiastic you sound about the rather obvious. As one of the top computer scientists, I find AI an ad-hoc for a decade not legit structural approach such as evo algos + decision trees + fuzzy logics. That would be the closest to ML, but orders of latitude more powerful, a story for a thousand years

[@underfitted]

I’m always enthusiastic 🤩

[@RazorflasK]

These are great takes from the best people of AI.

[@underfitted]

Yeah!

[@ochiojie]

I'm gonna be staying with your channel for years to come..W video! W editing! Amazing video king! 👑

[@underfitted]

Thanks, so much Ochi! Appreciate your comment!

[@keithprice1950]

I keep diving into data analysis and machine learning because I find them very interesting subjects. But then I always struggle to keep going because I hear so often that without a degree (I only graduated from high school) I will not be able to get a job in the field. This makes it hard to stay motivated. I wish there was a more definitive answer about getting into a machine learning role without a degree. I don't want to spend a couple of years learning and building only to find out I'm not employable.

[@underfitted]

You are employable as long as you know how to make things happen. Degrees don’t matter.

[@keithprice1950]

@@underfitted So a solid portfolio of projects and the ability to present them to a potential employer has the potential to get me an entry level position? I did think about getting my foot on the ladder with a data analyst position and then work up from there. Do you think that would be a better path?

[@RedShipsofSpainAgain]

What's the link to the Twitter thread where Rashka discussed these aspects of data vs hyperparameter tuning?

[@underfitted]

If you mean what I show in the video, that was a private convo. Happy to share it if you ping me on Twitter @svpino

[@sexyshortdude95]

@underfitted can you elaborate further what you mean when you said "a metric without context means very little" please.

[@underfitted]

Definitely! Imagine I tell you that I have a model that's 95% accurate. Without more context, that metric alone won't tell you much about the quality of my model. For example, if I'm trying to solve the MNIST dataset, 95% is very low. Or if humans can do the same task with 99% accuracy, 95% is not a great model. If my model solves a problem where humans are 80% accurate, then I have superhuman results and that's awesome! Metrics need context. Hopefully that makes sense.

[@aventurasenvideo]

Great quote "Since humans use ability X to perform Y and model does Y it must be through ability X" Will use this to teach my students the pitfalls of anthropomorphizing.

[@underfitted]

Definitely!

[@zekeluther]

Awesomeness

[@atharva_domale]

Doing Red hat system administrator course will beneficial for A.I background Student??

[@underfitted]

Every subject you can learn will somehow benefit you, but at least on the surface, there are courses with more direct benefit for AI than Red Hat System Administration.

[@raheemnasirudeen6394]

Nice

[@teja_surya]

surprised Bojan tunguz didn't include "xgboost" in his advice

[@underfitted]

Lol true

[@angadjabbal9211]

6:20 .. Causation doesn't equal correlation

[@canozturk369]

hey, ur channel is about to blow. just saying

[@underfitted]

Thanks!

[@machinelearning3518]

Reading Books and Courses is not enough, Solve a Problem, Every Tool has Limitations Don't see everything as Nail Don't Rigidly Follow any Learning Sequences Imposter Syndrome will be there Don't worry abt Academic Degrees Metrics without Context, Data Leaks, Use Features that are difficult to Collect Learn to Interpret and Evaluate ML Models Better Data > Hyperparameters

[@nguyengiahy6516]

Hi Santiago, can I use your message image with Bojan Tunguz please? I will credit your KZbin channel of course

[@underfitted]

Absolutely!

[@NavigatorlessName]

🎉

[@Atistatic]

The first mistake, is the same mistake like when people don’t practice math exercises enough.

[@DKtruster]

Good and informative video, but the speaking is a bit too loud

[@underfitted]

Yeah, you are right. I’ve been trying to fix it in my last few videos.

[@pipinstallyp]

Underrated video, but so much truth in it! Cringing at times when I said humans are ML systems. I thought I was being this mega intelligent human being but nah "I jUsT pOsTeDD CRINGE"

[@underfitted]

Thanks! Hope it helps!

[@KurtzMista]

Transformers are starting to look like silver bullets. What are your thoughts for beginners?

[@underfitted]

Don’t start there. Everything else is still very important.

[@alefratat4018]

No, it's definitely not.

[@MrAbhithepandey]

I am a machine learning professional, and 90% of my time is spent on cleaning+sourcing data. I did not got any new information from this video.

[@underfitted]

Thanks for taking time out of your busy schedule to let me know you didn't get any new information from this video.

[@vev]

😃

[@underfitted]

😁

[@robertocaropenaloza7764]

Control Theory Applications.

[@syzore2]

Top 1% don't have careers.. they have money.

[@underfitted]

That’s what they say

[@acsarda]

POR QUE GRITAS, POR QUE

[@underfitted]

Because they can’t hear me in the back!

[@acsarda]

@@underfitted haha touche

[@plazzmator]

hope selling

[@viktorbraa]

Why are you screaming bro😂

[@underfitted]

Because I want you to hear me, bro

[@alefratat4018]

The One Rule: Data > Learning metric > Model architecture > Hyperparameters tuning.

[@philsburydoboy]

Dude relax with the shouting...

[@underfitted]

Learning…

[@Carlos-bk2kp]

I will make your life easier, if you want to belong to the 1% of ML developers, learn math and be very good programmers.

[@shyiggy333]

1-8 having a life

[@underfitted]

Yes

[@maxmacken8859]

Bro, I love your videos, but please, you don't need to shout at us. Talk to us sincerely. The shouting just ruins it...

[@underfitted]

100%. Still learning. New videos are much better. Still room to grow

[@TheOzpad]

I like your topics but you're tying too hard to be Johnny Harris with your style of video. I can see you put a lot of work into editing and your studio looks great well done on the effort but for the content you are doing, your audience (assuming you are targeting people serious about ML) is going to feel lost becuase your videos feel like a Vice or Vox video, its edutainment. I'd like to watch the topics you have but without the jumpcuts, that is the number one thing that makes it hard to listen and watch your videos, its annoying. The shouting was touched on by somone else, they are right. Stop shouting please its hard to listen.

[@underfitted]

Thanks for the feedback. I’m trying to improve as much as I can as quickly as possible. This helps. Yes, these videos are different. I understand they aren’t what many AI/ML people are used to see. That’s okay. There are fantastic creators here on KZbin doing that type of content. I’m not one of them. I’m trying to bring a different voice (without shouting, hopefully). I’m trying to build a different audience. It’s okay if I turn of some people looking for something different. Thanks for taking the time to write your feedback!

[@TheOzpad]

@@underfitted Your channel has great promise and you will go far, try to get some cool guests on like Andrej Karpathy

[@underfitted]

Thanks, ozz! Absolutely! I'd love to have them on, and I will. Need to build some reputation first before they are willing to listen, but I'll make it happen, for sure.

[@rocking4joy]

why are you shouting

[@underfitted]

People in the back said they couldn’t hear me 😎

[@lcsmrt]

Great video!

[@underfitted]

Thanks!