'How neural networks learn' - Part I: Feature Visualization
Interpreting what neural networks are doing is a tricky problem.
In this video I dive into the approach of feature visualisation.
From simple neuron excitation to the Deep Visualisation Toolbox and the Google DeepDream project, let's open up the black box!
Links:
Distill.pub post on Feature Visualisation: distill.pub/2017/feature-visu...
Sander Dieleman post on music recommendation: benanne.github.io/2014/08/05/s...
Blogpost on Deep Feature visualisation: yosinski.com/deepvis
Github link to DeepVis Toolbox: github.com/yosinski/deep-visu...
Paper by Zeiler & Fergus: arxiv.org/abs/1311.2901
If you want to support this channel, here is my patreon link:
/ arxivinsights --- You are amazing!! ;)
If you have questions you would like to discuss with me personally, you can book a 1-on-1 video call through Pensight: pensight.com/x/xander-steenbr...
I watched this video when I wanted to fall asleep, to satisfy my conscience that I had done something in my assignment , and also hearing these subjects makes me fall asleep quickly, but your way of presenting the information was beautiful and not boring, so I could not sleep but became more active to complete my assignment ! Thank you & keep going ⚡️
Going through interesting topics with a direct link to research. You're truly doing amazing work.
Dude your content is amazing. So is Your way of talking and explaining... just amazing
Amazing videos! You have an awesome art of squeezing advanced concepts into a short, simple and interesting video!
This guy is super clear compared to everything else I have watched so far.
Fantastic channel! Great idea with highlighting the latest arXiv articles. I’ve been looking for a channel like yours! Subscribed and can’t wait to go through your content. Thanks for doing it.
Very nice video! Looking forward to the next parts. Love the dreamy images ^^
this is so cool that you're here to explain stuff. too many video are like "you copy this code, press play, boom, you built the 344th billion mnist image classifier, gg"
Really enjoying your videos, please keep it up!
This is really high quality stuff. I really appreciate the work you put in here. Keep it up!
Damn, this video is sooooo good! I subscribed to this channel after two minutes into the video. Great job!
i wan to mention that, your videos are super awesome! thank you!
oh man. this is wicked stuff! I really enjoy your style.
Very informative, and summarized knowledge. keep it up bro.
You are doing such a great job! I wanna see more :)
THANK YOOOOOOOU I was reading that article you commented on and I couldn't understand for the life of me how they were generating those images, so tysm ;-;
Excellent educational video on artificial and deep neural network learning.
Your channel is top quality. Thank you.
Using gradient descent to generate an image to max a neuron is cool, I suppose the same can be done with audio also? It would be interesting to hear what a neuron 'sounds like'.
I guess its the exact same process but you convert the generated spectrogram image into an audio file.
You are amazing! Please keep posting new content!
Great video! Cheers from Belgium,
you are really awesome brother. love from INDIA.
Quite dense, but vey well explained!
Excellent video, good job!
Awesome explanation!!
Really nice explanation :) Thank you
Thank you for making these videos
Really informative video, thanks :)
great presentation ! Thanks.
Nice video... explained pretty well..
Thank you! Good Explanation!
Bro! you were soo ahead of your time! Like Scooby Doo
Quality of your videos are the best. Good things take time but if you could upload weekly. It would be great.
Amazing video!
Amazing video👍🏼
Great overview! Although I'm not sure Zeiler and Fergus' work on feature visualization you've mentioned is actually training the deconvnet via backprob, I remembered that he mentioned that deconvnet is used as a probe to reconstruct image patches from the selected activation maps.
I agree with you. In the original deconvolutional network's paper (ieeexplore.ieee.org/abstract/document/5539957) they do train deconvolutional layers to reconstruct images in order to perform unsupervised training, but in the paper mentioned in the video they use transposed versions of the convolutional layers' filters to investigate a trained CNN for image classification.
Great video!
Excellent videos. I was wondering how can you evaluate and measure the layers inside, I've heard about these visualization methods so far. Thank you so much.
Great videos!
Your videos are amazing! Thank you
Can’t wait for part 2)
nice, i hope you make more videos. Good luck
Really good content
Awesome insights
Great video
This was just amazingggg
OMG feature visualisation LSD!
Hi Dear Arxiv, very good video! subscribed. any resource regarding how to look into a RNN （bi-GRU) to visualize the feature?
Thanks a lot, dude.
very nice video
This Thing can be used in Modern Art for sure. :)
thisartworkdoesnotexist.com
Super thanks for amazing videos! And I'm really waiting for the Part III :) You channel is the only one in my subscription list with the Bell turned ON :)
Super cool!
A trip to Bulgaria song :) Greetings from BG :)
Bravo!
The term "activation" in the context of neural networks generally refers to the output of a neuron, regardless of whether the network is recognizing a specific pattern. The activation is indeed a numerical value that represents the result of applying the neuron's activation function to the weighted sum of its inputs. Just posting here what ChatGPT told me, because the definition of "activation" in this video confused me
Nice video, carry up! :}
Here is a project that i created to focus on visualization. Hope someone finds it useful. Can be applied to images, audio, text or anything else: github.com/raghakot/keras-vis
Damn great video! Carry on ! This subject is so fundamental in Artificial Neural Networks : what the h*** do they learn ? ;)
I was here when this channel had 200 subscribers!
I wasn't
me neither
Feature visualisation is what an acid trip looks like
Great series of videos. The part 3 has been released??
Roxana Noelia Not yet, but I'm working on it! Hopefully somewhere next month :)
Thank you so much for doing these videos. This particularly one is great.
Best AI channel EVER
what motivated you to start this kind of videos? when and how you started? I love explaining what i know but: i am not as clear as you are, and i certainly do not have patience to learn all those editing tools
I wish there were more videos.
That tune is the bassi tune
At 4:20 you said the audio spectogram is converted to image. So how is it done? Like mapping those spectogram with image feature or those spectograms are somehow converted to an image.
I think deep visualization of games networks like checkers or go would be interesting.
Is the deconvolution a selective autoencoder?
How effective is this music recognition algorithm? Some www 2018 challenge showed the winners only got 60% accuracy - how can this be effective at such a low level?
When you say Neuron do you mean the filters of a convolutional neural network? Thanks for the video
Is there a way to visualize neural networks in general?
How visualizing works in playground tensorflow? What do they mean?
@Arxiv Insight - Cold start problem will not be solved by using deep nets to extract musical features. You will still not know what songs to recommend given a new user.
Also, there is a way to figure out if a neural net is looking at the sky or the ship. Black-out the sky and feed the ship - see the response!
which software are you using for video editing? plz reply...
how can I visualize the features in Matlab??
woow peacock in a educational video !
What if we try maximizing a certain class (e.g. dog) instead of a specific neuron?
you are so cool...
When's part 3 coming out?
Oh man, I really wanna make part 3 but I'm currently working on two episodes on Reinforcement Learning first, I wish I could do more episodes/month but currently I'm just too busy to work on this more than 10 hours/week. I need to find a way to increase my video output rate though :)
Is there any way we can help besides the Patreon? By the way, a fan of your contributions across mediums! ^_^ Really excited for the RL series -- I can't wait to be able to give back in the way you are once I've accumulated a bit of knowledge. How do you view this channel and its impact relative to any other work you're doing?
the c64 neutral net code did what these advanced nettoworks do back in the 80s in a single matrix input output layer reconstructing but with letters. it could not handle to much data as the same problem with it was the same as the modern version. the network get confused so make it big enough kind of solves it but not really. there is to many copies of mostly the same data in the network. im sure if such a basic linear function generated network like the 80s if big enough could feed a entire page of random letters and still make it reconstruct each trained letter correctly from random. think if you could do that with apples and bananas. im not sure if that would work but think in theory it chould. there is no reason why the network chould store the pattern of a banana or apple like a mess. i imagine the network can be trained with random objects and still reconstruct individual objects without classification.
With theory it is ok to explain, but in real world we need some code to implement it practically, so have some code, btw Ur videos are awesome
Has Part III ever been created?
Not yet, but I'm actually starting work on it right now. Should be finished in a couple of weeks!
@@ArxivInsights OK, great :)
Can neural network learn creat 3D enviroment from reading book specific chapter where author specificly is discribing that inviroment? Guys make this question viral!!!
why do those pictures look like an ayahuasca trip?
why you stop update new vedios ?
overall very good, but pops are a bit too loud and baby crying sounds at 1:28 are awful
But there should be some code
Faaaar better than siraj
noted
@@SirajRaval lmao
8:57 sus
Great content but unnecessarily long. You could say the same in half of the time. But really, that’s cool stuff 😎 congrats.
If Spotify is using a deep neural net, then why are it's recommendations still awful.........
Make sure you give it feedback. Really take some time to rate the songs in your recommended weekly, and after some time it'll get better!
test
Either stop waving your arms around, or focus more on your face and less on your hands .. its very distracting
Too much bla blaa blaaa blaaaaaaaaaaaaaaaaaaaaaaa...........................