MIT 6.S094: Convolutional Neural Networks for End-to-End Learning of the Driving Task

[@googlymoogly64]

We live in incredible times that these lectures are available online. This information is crucial for my work and I have no idea how I'd be able to educate myself if I didn't have access to these lectures. I really appreciate your work.

[@AndreiBarsan]

I completely agree! Stanford also have TONS of ML-related lectures (+notes and other materials) available online. Ian Goodfellow's DL book is also readable online, and I've heard quite good things about Udacity's ML and DL courses. (They even have a course on self-driving cars, but it's not accessible for free IIRC.)

[@abcxyz4207]

Without a degree it’s worthless, that’s why they upload it

[@leroyjordy9012]

i know it's kinda randomly asking but do anybody know of a good website to watch new movies online ?

[@remyelliot1264]

@Leroy Jordy Try flixzone. Just google for it =)

[@TheMicahwitz]

Isn’t it insane??? I can’t believe how many I’ve league classes you can watch online.

[@tazojj3717]

I am 90 years old, still read 2 books a year. Love your mind people!

[@IFlyIChris]

🙏🏾

[@Bojonni]

If you don't use it you'll lose it

[@helloworld4788]

Cap

[@recursion.]

@@helloworld4788fr people still falls for it.

[@manish8206]

U r young, young means who wants to learn.

[@nadjabekele9280]

For those who already have a background in CNNs, the video starts at 36:15

[@IndrajitRajtilak]

Thank you!

[@ArchiRuban]

You're a legend

[@tellmebaby183]

you are a keeper of heart

[@zachaIIen]

I can't thank you enough for uploading these lectures.

[@TheMicahwitz]

Lex is such a boss. The more I learn about him the more impressed I am.

[@bocckoka]

Lex teaching the people the correct pronunciation of GIF, very nice.

[@SF-fb6lv]

0:43 as soon as Lex said "...the problem with detecting green, yellow, red..." I realized my project in ML was not trivial...

[@ahmetfarukcakmak6804]

Absolute perfection. Thank you Lex for all lectures.

[@thegoatmen7736]

Lex does a great job explaining complex problems in a way which even uneducated folks such as I can even understand.

[@truthforall6651]

Lex is a funny teacher 😂😂😂

[@sumeahsking8019]

Thank you for making education more available with no monetary cost

[@ProfParzival]

You’ve come a long way Lex, proud of you man

[@double_j3867]

This guy seems way more like an engineer than a com sci background. That's a compliment (and a poke at CS majors) :-)

[@mrdgg949]

This guy should start a podcast

[@udoyxyz]

Wow. You look great

[@MattCamp]

I really like your lectures! Thanks for posting them!!

[@MikeFizzyD]

Your lectures are nice and straight to the point/application. Thank you

[@ShravanKumar147]

Thank you very much for your great lectures, could you make a playlist with numbering in order.

[@lexfridman]

Thanks! Here's a playlist with the lectures uploaded so far: kzbin.info/aero/PLrAXtmErZgOeiKm4sgNOknGvNjby9efdf

[@murk959]

I follow him on LinkedIn and just realized he's an actual professor despite his bio saying he is

[@philwilson1445]

Thanks. The work by NVIDIA is so amazing. Will you be uploading the guest lectures too?

[@lexfridman]

Thanks! I'll be uploading all the course lectures and the guest talks as well.

[@philwilson1445]

That's Great. Already waiting for more material !!

[@svenhofstede]

Thank you so much for this Lex

[@srebrnimedved]

The tone is like we're being trained in bjj

[@wiredog771]

Absence of black tie does not compute.

[@carlosespinosa6446]

THANK YOU! Computer Scientists and more people will appreciate :)

[@annunakian8054]

Lex's ivy league body count must be immense.

[@loordvamp1re]

Would like to hear more of similar stuff from you ^^ really great :D

[@mahneh7121]

Wow, i just learned about NNs from Andrew Ng. Now this lecture was fantastic as well. So we could think that our neurons also have this weights and they fire depending on some condition, or abstract aspect in the query image ? Thank you so much man. A random chemist here 🙂

[@eaf888]

come a long way!!

[@triplef3v3r]

Hey Lex I sent you an email for today's class, please open it and read it in the classroom. Thank you. Oh btw, good job doing interviews!

[@markjing9124]

Thank you !!!

[@microdrone]

Lots of great info on the nature of the subject of ai/nn in general!

[@pastas6109]

I absolutely love this! Thank you so much!!!

[@therealjeffstewart]

Around 52 min I could imagine space travel at incredibly high speeds. Like the beginnings of light speed or plaid? Also, the LIDAR reminds me of images of V for Vendettas surveillance vans.. V for Vendetta depicted a world of total surveillance. Government agents roamed the streets in spy vans, listening in on everyone from average citizens to the Bishop of Westminster, in order to seek out dissenters and other threats to those in power - all in the name of public safety. Nov 5, 2013 But now every government agency will have these driving around. Hope this all remains friendly.

[@techonlyguo9788]

The videos in this video is amazing !! Thanks !!

[@nosferatutheun-dead6433]

I won't mind flunking my paper every now and then if that's what it would take for me to be forever in his class :P

[@igorszemela1610]

thanks a lot for sharing

[@lanyastar]

A short question: when traning a cnn network with back propogation , how the parameters change when passing through the pooling layer?

[@陈典-v2m]

Thank you for the excellent lectures! They are fantastic! And are the guest talks available on youtube (can't find links on the course page) ?

[@lexfridman]

I'm working on making those available soon. Sorry for the delays.

[@AndreiBarsan]

Thank you VERY much!

[@balabuyew]

The point I can't understand intuitively - is why every "filter" in convolution layer is a single neuron. Why its not a network (of several neurons) in general case?

[@nguyenthanhdat93]

You can imagine each "filter" only detect certain parts of an image (a wheel, a window, a license plate). The deeper the filter in ConvNet the more general it could detect based on previous filters. For example, the first few conv layers near the Input Layer detect wheels, license plates, color, edges. The last few layers before the output can "filters" cars, truck, police car, etc. I think the word "filter" implies that a convolutional layer produces an output or being stimulated when it detects a part in an image. A "wheel" filter produces an output if input has a wheel. The stronger its confidence about the wheel in an image, the bigger the output (or probability).

[@balabuyew]

But, the filter should be able to extract a feature to propagate it to the following layer. What if a feature is not so simple to extract by single neuron.

[@nguyenthanhdat93]

If I understand your question correctly, your concern is how can a filter could "see" a complex feature. I think that is why they suggest to build a "deep" ConvNet so that I could detect more complex objects. For example, ResNet from Microsoft, a winner from ImageNet in 2015, is a very "deep" ConvNet.

[@balabuyew]

The question is whether these two architectures (deep conv nex with single neuron filters, and, less deep conv net with more complex filters) have equal expressive power.

[@salehabushatara6447]

Lex iOS 10

[@md.shahidulislam3178]

Your lectures are really grate (Y)

[@yoniami3665]

Thank you so much! is there any way to see the tutorials or the slides of the tutorials?

[@theguyman232323]

Interesting. I had to use tensorflow to identify outputs for those exact same 28x28 pixel images @21:09 of numbers at George mason. I guess that data just gets circulated between colleges

[@SageBetko]

This is a canonical image classification dataset, MNIST. Anyone can use it!

[@IsItEvenWorthIt]

This guy will never be a podcaster

[@arielalaniz7819]

LOL look at you look at him now mf

[@IsItEvenWorthIt]

@@arielalaniz7819 it’s a joke stupid.

[@SubmitTheKraken]

This is just unbelievable, great vid! Is there the possibility that you would give a lecture about computational neuroscience as in the atom memristor etc.

[@Snowflake_tv]

You were a professor? Wow...

[@TamasKalman]

i am surprised tesla didn’t build an insane amount of learning data just by recording all camera on all cars 24/7 and send the video stream to their datacenters along with gps data + driver inputs + all sensors in the car load it into a multipeta database all roads in all conditions and times recorded by thousands and thousand times i’d assume a convolutional network would learn something out of it by now how to drive those roads aren’t the biorobots are used to train the ai which will replace them?

[@inquisitortr7930]

My final project at school is autonomous driving with CNN. I have a problem with softmax function. I should take 3 steering output (right, left, straight) at the end of the CNN however how should CNN calculate steering angle values. Right,letf,straight do not mean anything if i dont know how much steering value. My training set outputs (steering values) changing between -1 and 1. Note: Softmax function do not accept float values. It just accept integer and booleen.

[@kinvert]

This guy could become a big deal some day.

[@sandragarcia-fk5vi]

compatible version of jupyter notebook for tensorflow

[@endgamefond]

I like days when lex wearing eyeglasses

[@demonicode297]

please upload more lectures . tnx

[@RobetPaulG]

1_python_perceptron.ipynb has an xrange() method call. Stackoverflow says this method isn't available in Python 3 which is necessary for tensorflow. Use range() instead.

[@RoscoeDaMule]

lol I had no clue lex was like a prof. or whatever this sheet is

[@gk7001]

So I got as far as the first slide gave up. How far’d you git?

[@kylint7683]

If I remember correctly, there is also a rich kid got killed in Tesla's autopilot mode while driving in China. So that's 2 fatalities per 3million miles

[@masbro1901]

what major is this, is it electrical engineering or informatic engineering or electronic engineering

[@EpicMathTime]

Computer science, mathematics.

[@nyquist_control]

None of those 😂

[@25sigmaa]

🥰🥰🥰🥰

[@pravachanpatra4012]

1:02:07

[@viralgupta5630]

Are there good examples that explain step by step Image Segmentation and Object Detection. Its easy to find examples of image classification.

[@AndreiBarsan]

There's quite a lot of work on that, though, I agree, it's not as popular as the "vanilla" classification. The "oldschool" way for object detection was based on sliding window approaches + something called "object proposals". Newer technologies for object detection include (Fast/Faster) R-CNN. As far as segmentation goes, there's also quite a lot going on. Try looking up SegNet, Multi-Task Network Cascades, FuseNet, ENet (this one is new, real-time, AND available on GitHub!). For a gentle introduction (which, granted, doesn't go into deep learning that much, but is still very useful), I'd recommend the Udacity CV course. Hope this helps!

[@frankdinies5111]

Maybe you find this usefull: www.coursera.org/learn/convolutional-neural-networks

[@benji1570]

Is your brother playing in Denver Nuggets 🤣

[@umeshkacha1638]

Thanks much what is that software which gives real time detail of face in a video shown at 1:04: 23?

[@coincrazy3563]

welp found what in watching today

[@shahaan-shah]

Why is he dressed so casually

[@nyquist_control]

Is that what you're seriously thinking about?

[@phaedruslykos3249]

one like because computer vision is hard

[@SuperMaDBrothers]

This is a pretty poor explanation of this topic. Someone seeing it for the first time would have no idea what is going on. All the material is there though.

[@thalberg-]

I don’t understand why he is lecturing like he got forced to take over an absent colleague/is slightly high and struggling to get it together? The orders of slides/presentation details are super weird at times too, not logical in order???

[@jagrcarl]

Lex is a known coke user... ;)

[@tejasmic]

Carssss

[@valeridonkov7951]

And Luis Hamilton still will beat you with his eyes closed, it's embarrassing how behind you are!

[@danielgray8053]

I don't like that he has to look at his notes every 3 or 4 words. Is he reading a pre written speech? I like professors who are just spewing knowledge form their brain.

[@go00o87]

How can a conv. neural network perform better at classification than humens, if humans label the images? or is the labeling done in a different way?

[@lexfridman]

Yes annotation for the ground truth is done differently than how the classification task is performed. The former is a binary verification of an already labeled image (from Google Image Search). The latter is pure full categorization task. Human performance as a classifier isn't a perfect measure in this case. Read more about it here: karpathy.github.io/2014/09/02/what-i-learned-from-competing-against-a-convnet-on-imagenet/

[@go00o87]

Thank you vary much for the link and the quick response. If I understand it correctly every image gets one label which is then referred to as “ground truth”. I agree it is a useful concept up to a certain point. Because ground truth is a function with the input value being an image and its out put value is a label. And there is no unique way to do this. Now you ask humans and NN to approximate this function. It is impressive that they perform roughly the same. But the insight in performing better than humans is limited. Just Imagen what it would mean if an NN could achieve 100%.

[@lexfridman]

Yes. Exactly. Your observation is spot on. Still it's a useful test, but there should be many others.

[@timelessone23]

Jordan Peterson mentioned somewhere that we do not see things, but whst you can do with it. That is, you do not see a door, but a 'going through' thing. So in a way, the labeling might need to include that too?

[@jakefloden6512]

Lex in jeans, holy fuck he is human

[@vidius-lordbrandonthedarki9027]

Is this peer reviewed?

[@heygema]

LMAO

[@csankar69]

Horrible lectures with no real content from which one can benefit. Feels like a summary presentation for a set of journalists who don't want to get technical!

[@ryandsouza9042]

Hmmm, It would be so nice if you could take your comment and Fuck Yourself! While the rest of humanity benefits from these lectures,you should seriously consider my advice and Go Fuck Yourself!

[@thedeathstar420]

@@ryandsouza9042 Sorry for the late reply but sadly I aagree with csankar69. This is not helping humanity in any shape or form because there are numerous lectures just like this all over youtube. People who actually want to have a strong technical understanding in CNN are starving for more rigor and this dude is using MIT resources to make summary lectures.

[@dco1019]

@@thedeathstar420 from what i know this is fridmans own work. Or otherwise said this is the kinda work he did or still does.. So he could certainly go into the nuts and bolts. But, correct me if I'm wrong, the nuts and bolts are probably lines upon lines of code to make machines do very simple things. In my estimation that wouldn't do much for humanity either in the sense of planting seeds and motivating the youth to take interest.