Thomas Pinetz

@RokoMijicUK @goodfellow_ian Something you should also not underestimate are good libraries for DL.

@goodfellow_ian > barriers to the success of deep learning before 2012 were largely psychological What about the issue of needing more compute power? At what point did this cease to be a constraint?

For years I’ve believed in the story that after the first person ran a 4 minute mile, suddenly many other athletes found that they could too, and that the 4 minute barrier was largely psychological. Today I found this article arguing that’s a myth: scienceofrunning.com/2017/05/the-ro…

@avdnoord @catamorphist @Vinyals On how many GPU's for how long was this trained for the imagenet 256x256 images?

VQVAE-2 finally out! Powerful autoregressive models in a hierarchical compressed latent space. No modes were collapsed in the creation of these samples ;) Arixv: arxiv.org/abs/1906.00446 With @catamorphist and @vinyals More samples and details 👇 [thread]

@MioMilenkovic @goodfellow_ian It is probably a combination of the dataset (showing celebrities which on average are quite good looking) and cherry picked results.

These style-based generator results look great: arxiv.org/pdf/1812.04948…

@fhuszar Even on images you have to care for your data. Adjusting the images to a useful range of values, using useful augmentation strategies or adjusting the resolution have all increased the performance of my CNNs.

Unpopular myth-busting opinion: deep learning DOES NOT do away with feature engineering. In certain high-D dense domains such as images or sounds, convolution-like things do well on what we call “raw” data. Elsewhere, input representation matters. Let the flame wars commence.

@daniels_kaiser @xLaszlo @ylecun @sebastiankurz That is actually quite impressive given the size of the country.

Many European country should dramatically increase their R&D investments. howmuch.net/articles/resea…

@fchollet Can we also have examples for this or do we just condemn communities on prejudice nowadays? Racist/Sexist comments are downvoted into oblivion, as has been shown in practically all recent threads.

If you are ever in doubt that we need to keep doing more to make the ML community more welcoming and more diverse, you can always look at the ML subreddit, where, whenever the topic of diversity comes up, the must upvoted comments are always the most disgusting and toxic.

@fchollet But even they show that the progress was real. The performance has been increasing linearly since the introduction of AlexNet, even if the final accuracy is not as good as promised, which might also be due to a harder testset.

Stating the obvious: a lot of current deep learning tricks are overfit to the validation sets of well-known benchmarks, including CIFAR10. It's nice to see this quantified. This has been a problem with ImageNet since at least 2015. arxiv.org/abs/1806.00451

@beenwrekt I expected there to be a bigger change in accuracy. My main takeaway is that there has been significant progress in DL and improvement rates scale linearly with your new dataset without any severe outliers.

Do CIFAR-10 classifiers generalize to CIFAR-10? The answer will surprise you! arxiv.org/abs/1806.00451

@gstsdn But why though? The Wasserstein formulation requires a near optimal critic function in their mathematical justification. Training for a single iteration seems insufficient. How to train better generative NN is cool, but your paper does not improve performance in general on WGANs.

@samonlinestores I ordered a whiteboard on the 18.12.2017 and I think there are some problems with the delivery. I contacted PostNL, but they did not know my local sender. Would you be so kind and solve this problem? Order Number: 84003502.

@PostNL Could you tell me who the sender is? I can not find the information anywhere.

@PostNL According to your website I have missed the delivery of my parcel 3STAQT1988269 (in Austria) . What should I do? It has been a week and the driver did not leave any notification..

@porestar @_AntreasAntonio @jeremyphoward @yaringal @soumithchintala In the original Improved Wasserstein paper (arxiv.org/pdf/1704.00028…) they found that two sided gradient penalty converges faster to better optima, than one-sided penalty (wgan-lp).