GPT-3 Roundup Discussion and Demos

 Here it is in all it's glory.  Hot off the presses (very recently posted).  An extended panel discussion concerning GPT-3.  Where learned individuals will discourse.  One of them (Gary Marcus) is a notorious deep learning curmudgeon, so be warned.  The others are all pretty enthusiastic (the moderator a little bit less than some of the others, and the linguist is a linguist so he's always going to have issues).  It's all quite fascinating.

The beginning is kind of a fast paced, somewhat chaotic overview of key insights discussed at different points in a somewhat long series of different discussions, and then the rest of the video is the actual discussions and some experiments with GPT-3 by the moderator.


Observations

1:  Gary said something about DeepMind's Atari game system that was very misleading.  He points out that if you move the location of the game paddle, then the trained system is no longer able to play the game.  Implying this brittle-ness, and that the system didn't really learn how to play the game.

I see this as just a data augmentation issue.  And it's an important distinction that Gary overlooks (probably because he has his own agenda in trying to rip apart the performance of deep learning systems).

Think of an image classification system. So if you naively train it on images of dogs and cats, and then you horizontally or vertically flip the input image, and then the trained system no longer works as well.  A direct analogy to what Gary is talking about.

Sure, the system learns what you trained it on.  If you introduce data augmentation in the training (randomly flipping and otherwise distorting the input), then the system will learn the correct transformation.  And it now works on thinks like flipped, or distorted input images.


2:  Conner briefly mentions manifold learning and latent spaces as another way to think of how deep learning systems work.  This is the key concept in my mind.

I was listening to an interview of Ian Goodfellow by Lex Fridman, and Ian made some comment about how his thinking on neural nets had evolved and he thought of them as learning a series of programming steps, and i remember thinking (with all respect to Ian) that was disappointing that he thought of it that way, because you really want to think of these systems as learning a low dimensional manifold mapping of the data.  

And why does this work.  Because data about the real world actually lives in lower dimensional manifolds (lower compared to the total variability of the components of the signal).  The neural net is a universal math function approximator.  And the function that it is learning is this manifold mapping function.

Even though they discuss manifold learning, 5 minutes later they drop back into the misleading 'it's computing a program' mindset.


Comments

Post a Comment

Popular posts from this blog

CycleGAN: a GAN architecture for learning unpaired image to image transformations

Image
 Like the Doublemint Twins touting the joys of Doublemint gum, 2 GANs are surely better than 1 GAN.  Especially if we package them together inside of one meta GAN module.  And this is exactly what the CycleGAN architecture does. Have you ever harbored dark secrets of turning a horse into a zebra?  The CycleGAN was developed to do just that. Learn how to turn a horse into a zebra.  And more. Now right away you can notice a difference between the image to image transformation GAN architectures we've been discussing over the last few posts.  Those last few posts described systems that learn from a database of matched input-output image pairs. And if your goal is to turn an edge representation into a nicely filled in continuous tone image, it's easy to build your database of matched input-output image pairs that your GAN system can then learn off of.  Take a continuous tone image (which will be the output of the database pair entry), then run it through an edge detector algorithm. 
Read more

Pix2Pix: a GAN architecture for image to image transformation

Image
 I thought following up yesterday's TraVelGAN post with a Pix2Pix GAN post would be useful to compare what is going on in the 2 architectures.  Two different approaches to the same problem. I stole this Pix2Pix Overview slide below from an excellent deeplearning.ai GAN course (note that they borrowed it from the original paper) because it gives you a good feel for what is going on inside of the Pix2Pix architecture.    Note how the Generator part is very much like an auto-encoder architecture, but rebuilt using the U-Net architecture features (based on skip-connections) that fastai has been discussing in their courses for several years before it became more widely known to the deep learning community at large  (and which originally came from an obscure medical image segmentation paper) . So the Generator in this Pix2Pix GAN is really pretty sophisticated, consisting of a whole image to image auto-encoder network with U-Net skip connections to generate better image quality at highe
Read more

Smart Fabrics

Image
"What makes smart fabrics revolutionary is that they have the ability to do many things that traditional fabrics cannot, including communicate, transform, conduct energy and even grow."  This is a quote from Rebeccah Pailes-Friedman, who is a professor at Pratt with a research focus on developing wearable technology and smart textiles. Enhancing textiles can be viewed through different lens.  Enhancements can be aesthetic, or performance boosting, or durability boosting. 
Read more