- 22
- 88 889
David Bau
Приєднався 3 лис 2007
David Bau PhD Defense
Dissection of Deep Neural Networks - David Bau's PhD Defense at MIT, August 24 2021
Do deep networks contain concepts?
One of the great challenges of neural networks is to understand how they work. Because a deep network is trained by an optimizer, we cannot ask a programmer to explain the reasons for the specific computations that it happens to do. So we have traditionally focused on testing a network's external behavior, ignorant of insights or flaws that may hide within the black box.
But what if we could ask the network itself what it is thinking? Inspired by classical neuroscience research on biological brains, I introduce methods to directly probe the internal structure of a deep convolutional neural network by testing the activity of individual neurons and their interactions.
Beginning with the simple proposal that an individual neuron might represent one internal concept, we investigate the possibility of reading human-understandable concepts within a deep network in a concrete, quantitative way: Which neurons? Which concepts? What role do concept neurons play? And can we see rules governing relationships between concepts?
Following this inquiry within state-of-the-art models in computer vision leads us to insights about the computational structure of those deep networks that enable several new applications, including "GAN Paint" semantic manipulation of objects in an image; understanding of the sparse logic of a classifier; and quick, selective editing of generalizable rules within a fully trained StyleGAN network.
In the talk, we challenge the notion that the internal calculations of a neural network must be hopelessly opaque. Instead, we strive to tear back the curtain and chart a path through the detailed structure of a deep network by which we can begin to understand its logic.
Dissertation, slides, demos, code, and data at dissection.csail.mit.edu/
Do deep networks contain concepts?
One of the great challenges of neural networks is to understand how they work. Because a deep network is trained by an optimizer, we cannot ask a programmer to explain the reasons for the specific computations that it happens to do. So we have traditionally focused on testing a network's external behavior, ignorant of insights or flaws that may hide within the black box.
But what if we could ask the network itself what it is thinking? Inspired by classical neuroscience research on biological brains, I introduce methods to directly probe the internal structure of a deep convolutional neural network by testing the activity of individual neurons and their interactions.
Beginning with the simple proposal that an individual neuron might represent one internal concept, we investigate the possibility of reading human-understandable concepts within a deep network in a concrete, quantitative way: Which neurons? Which concepts? What role do concept neurons play? And can we see rules governing relationships between concepts?
Following this inquiry within state-of-the-art models in computer vision leads us to insights about the computational structure of those deep networks that enable several new applications, including "GAN Paint" semantic manipulation of objects in an image; understanding of the sparse logic of a classifier; and quick, selective editing of generalizable rules within a fully trained StyleGAN network.
In the talk, we challenge the notion that the internal calculations of a neural network must be hopelessly opaque. Instead, we strive to tear back the curtain and chart a path through the detailed structure of a deep network by which we can begin to understand its logic.
Dissertation, slides, demos, code, and data at dissection.csail.mit.edu/
Переглядів: 17 618
Відео
Rewriting a Deep Generative Model (ECCV 2020)
Переглядів 3,9 тис.4 роки тому
Rewriting a Deep Generative Network, presented at ECCV 2020 (oral). In this paper, we show how a deep generative network can be reprogrammed by directly rewriting model weights. rewriting.csail.mit.edu github.com/davidbau/rewriting arxiv.org/pdf/2007.15646.pdf Rewriting a model is challenging, because doing it effectively requires the user to develop a causally correct understanding of the stru...
Rewriting a Deep Generative Model (Preview)
Переглядів 2,2 тис.4 роки тому
You wouldn't normally fiddle with weights of a state-of-the-art deep network after weeks of training. But why not? In this ECCV 2020 paper, we show how the rules of a deep generative network can be changed by directly editing model weights. rewriting.csail.mit.edu/ github.com/davidbau/rewriting/ arxiv.org/pdf/2007.15646.pdf Rewriting a model is challenging, because doing it effectively requires...
BEG practice
Переглядів 2838 років тому
My total beginner guitar "brown eyed girl" practice - an exercise in learning something totally new in a week.
Pencil Code - Angles and Arcs
Переглядів 9 тис.11 років тому
Programming and geometry with Pencil Code: this video talks about exterior and interior angles, regular polygons and arcs.
Functions for Drawing in TurtleBits
Переглядів 21611 років тому
Drawing Lines, Angles, Curves, Dots, and Text in TurtleBits.
TurtleBits
Переглядів 67511 років тому
A first look at TurtleBits, a new tool for learning how to program. Visit turtlebits.net/ to give it a try.
Blue Gear Ticks at WPI Robonautica
Переглядів 68012 років тому
Catherine and Dante run the Blue Gear Ticks robot at the First Lego League Massachusetts Championships 2011.
Blue Gear Ticks at WPI
Переглядів 27712 років тому
Anthony and Thom run the Blue Gear Ticks Robot at the 2011 WPI Robonautica First Lego League Massachusetts Championship.
Raging Hair
Переглядів 9313 років тому
Meet Pinky, the superhero with Raging Hair Power. This is our family's first experience with our GS150 and iMovie. The children like to come up with superheroes who have unusual super-powers. In this episode, you get to meet Pinky. She is a very nice girl, but you don't want to make her mad. Piper and Anthony and Dad Playing Herself: Pinky
Nonahexaflexagon
Переглядів 2,2 тис.13 років тому
This is a Tuckerman's Traverse of a Nonahexaflexagon. A hexaflexagon is a strip of paper folded origami-style into a flat hexagon in a way that allows it to be flexed to reveal additional hidden faces. A nona-hexaflexagon has nine faces total; Flexagons were first studied by Stone, Tukey, Tuckerman, and Feynman, and popularized by Martin Gardner. David Bau
Turbot
Переглядів 1,2 тис.13 років тому
Here is our Turbot. The Turbot is a very cool tumbling light-seeking robot that you can build from a Solarbotics kit. It has two long legs that it can use to crawl around and even climb over obstacles.
Piper Creates a New State of Matter
Переглядів 2,4 тис.14 років тому
Piper Creates a New State of Matter
The example at 12:14 , I’m curious if you could have used more descriptive language like “Only change the bed to be green while maintaining tan walls”? To be clear though, I agree it is a problem for a generator like this when you do not understand how it works and are trying to control it this way. Good work!
We need more videos from you! If you have lectures available and are comfortable sharing them please do so.
Wow, Dr. Bau is really pedagogical and is pleasant to listen to.
Nice work. Bushy eyebrow kids isn't something I thought I would hear today 😅
your work is so awesome
This presentation has blown my mind. Thank you for making this publicly available!
Dear prof Dr sir Thank you very 🙏 much 🙏
Really cool results !
nice. That D shape on the top can be fretted without removing you fingers after the G and C. Use your little finger to do the top string on the G and then use your second finger for the 2nd fret top string of the D. in that way you never have to move your hand.
Great job! I really appreciate what you have achieved.
Can you help me, I want to make one for my son
Amazing! A longer video could add so much value
Summa cum laude?
the video has been greatly intriguing, and really enjoyed watching it. thanks for sharing 58:16
Thanks for sharing, it was a great watch! Can you please comment on: 1. How do you search for "concept neurons" in giant models efficiently? 2. You mentioned in another comment that your group is working on interpreting GPT-X models, can you briefly comment on the concepts your group is trying to find out in GPT-X based models?
"PhD Defense at MIT" and for most people, he is "the amazing maze guy". Due to pandemics and KDP explosion, his mazes are now in demand. A lot of people is trying to figure out how to make solutions script.
Check out this link - web.mit.edu/PostScript/obfuscated-1993/labyrinth.ps - it is my submission to the obfuscated postscript contest web.mit.edu/PostScript/obfuscated-1993/WINNERS and it generates random mazes together with a solution, with all the computation done on the postscript printer. Edit the postscript as a textfile to change the options.
"You need to hold on tight to your optimism. Because if you don't, You're never going to figure out how to do the hard things to answer the hard questions and make it really work"
I am at 27 minutes mark and I am impatient to ask this question with out completing this video that does this method scale up to models which have millions and billions of parameters
Yes. We have found that the largest models, trained for a long time on massive data sets, tend to have very rich interpretable structure. Developing interpretable methods for massively parameterized models such as GPT-X is the topic of ongoing work in my group, and we have found that large models are a very target-rich environment. Oddly enough, one of the more difficult problems is to clarify is how interpretable structure emerges in the very simplest toy settings, trained on small problems, where the emergent structure is less obvious.
Congrats! This was really interesting!
Congratulations Dr. Bau! It is a great and systematic research project and a very important one with many practical implications. I am sure it could be a great step towards a transparent general AI if you start combining and interconnecting such separate networks together at the interaction levels. I am fascinated! Thank you very much! It really inspiring!
Great watch. Thanks for sharing this, David. - Ervin
By the way, if you or somebody you know is considering a PhD, I am looking for students! (For Fall 2022.) Check out our papers davidbau.com/research/, apply to the Khoury school www.khoury.northeastern.edu/apply/phd-apply/, and drop me a note if you are interested.
Dr. Bau , what is the best way to reach you ?
Congratulations!
Thank you!
I can't join the chat without creating a channel, which I don't have time to do - but I wanted to drop in and congratulate you, Dr Bau!!
Thank you Kate!!!
awesome cannot imagine how many new memes created if adobe put this on photoshop ; )
Awesome bro
Loving the vids!! Keep the videos coming! I think you should search for smzeus . c o m. It will help you promote your videos.
Absolutely brilliant and wonderful. Well done!
This is my favourite paper of 2020. Amzing work!
incredible
I need help I tried to use your code at 1;57 but it doesn't work
Whats the deal with silicon valley? I keep hearing video games from that place.
This is one of the best videos of the Solarbotics Turbot on the web! It's so cool how it got around the obstacle at 0:50
Wow! Thanks!
фуууу
How du u make it
are you guys going to make an app for iPhone and Android, that would be very cool😍
Hi David - great video! When I try the code, I don't get the exterior angle degree measures to draw (I do see the orange line extension and the orange arc). Any ideas why that is happening?
Matthew Fahrenbacher - there is a special script that you need to load to get the orange markers. Add the following to the beginning of your program: await loadscript '/lib/angles.cs', defer()
David Bau Thanks for getting back to me. I had already included that loadscript command in my program - it does cause the orange arc and lines to appear, but NOT the degree measurers. Here is my program: matfah.pencilcode.net/edit/ExteriorAngles
Ah hah, I understand now. That's a bug in angles.cs! I have fixed that library now.
David Bau Thanks! It works perfectly now.
I'm confused as to what's happening here...
My team was there as well.
Thank you Piper! Best wishes. I am amazed by the game at 1:57!
Yeah, that program is a ton of fun. You can play with the orbit program here: guide.pencilcode.net/edit/motion/orbit
Thank you David. A couple of years ago, me my 6th grade daughter spent at least 2 days in excel spreadsheets and later python pygame scripting to calculate orbits first using sin.cos formulae and the using newtonian equation. She learned a lot. But none of those frameworks can come anywhere near what you have now. With your platform, we can probably do that in an hour, focusing right only on mechanics of orbits rather than the mechanics of the language. Awesome.
Please pardon me but i can not help shower more praise to you - at the elegance of your expression (code), not one unnecessary letter! what a masterly art. I have previously learnt from your blog (RPS, your js tutorials) and have a great appreciation for your verbal computational and visual imagination and generousity. Thanks for sharing the great work. I will introduce pencilcode to more kids.
Ravi Annaswamy Thanks for the kind words! Think about hosting a "CS Education Week" group December 9, e.g., masstlcef.org/hour-of-code/ . The idea is to denote a week to inspire kids to learn CS by providing a one-hour easy-to-teach CS lesson. Pencil Code "Hour" materials are being posted here: event.pencilcode.net/hoc2013
thank you, I will consider, that is a great idea
David, You have created an amazing platform! I am sure this will catch up really big! It will add so much value to kids! BTW the tag chase and orbit just blows my mind. You have such a way to minimally do things!
agree even though i am a kid
Interesting turtle bits
Note that the site is now called "Pencil Code" - I have just published a short book to go along with the site: pencilcode.net/wish
Wow! Sheer Genius, David!
Just a guess, the characteristic dimensions of both beads are such they establish a resonance beats that are offset from one another. A way to check would be to step the frequencies around until resonance peaks are observed for each bead size then compare them to the wavelength... very cool! Dad you are doing something right, especially if she picks the reins up a little more in the future... Luck!
So how do u make it ?
Man, that kid is HARSH! did you make it rechargeable?
oh yeah she totally did it by her self, cheater!!