Видео

Heyyy can you please provide source code or repository??

well done! guys. Can I ask some questions about the method you used?

awesome

Amazing work!

Very interesting tutorial. Great work!

Great paper. Any intuition on why the ViT Lite learns better with learnable PEs compared to sinusoidal?

Is this paper accepted by CVPR 2021?

This was a really good talk. In fact, one of the best explanations on entropy minimization. Thank you very much for your good work.

Thanks! I was so confused about why entropy minimization can push the decision boundary

Nice. You are welcome to watch my playlist : ruclips.net/p/PLdkryDe59y4Ze9_12JhWu3cs-lOGYwYeD Eran

Problem with audio between 12:49 and 13:45. Would it be possible to upload a transcription for this part? Thanks.

Thanks for the talk and for sharing the Toronto Annotation Suite. Is there any reason aside from publication time/project timing that StyleGAN was used instead of StyleGAN2?

Very informative! Thanks

plots not starting at 0 on the y axis + no error bars make it quite hard to actually compare performance...

Thanks for the nice talk. Your idea of weighting unlabeled data points in the semi-supervised context is certainly related to learning from "soft" probability distributions as in ReMixMatch. There, you also have some sort of data weighting by potentially less extreme pseudo target distributions (as opposed to hard labeling as for instance in FixMatch). That being said, I think many semi-supervised pseudo-labeling approaches implicitly implement such kind of instance weighting. Nevertheless, I like seeing that you are treating this in a more explicit yet "classical weighting fashion". Great work!

Thanks for the talk. I have a couple questions what other methods are there to promote flat minima and why what this method chosen? How much slower does this make training since there is now a adversarial optimization program to solve?

Hi I really enjoyed the talk. I have two questions, do the transformers introduced have a similar ability as current visual transformers to improve performance with dataset size or do they saturate at some point? Secondly, it appears a core design pattern of this model is combining transformers with convolution what is the right tradeoff here? Do you think there might be some way to control the tradeoff between low bias(transformers) and higher bias(convolutoion) in a way to enable a "smooth" tradeoff?

Thanks for the nice talk. Is there an intuition about why sequence pooling improves performance? It looks like its learned weighting is not dependent on the input. Does this pooling learn some bias like always giving a larger weight to the patches that came from the center of the image?

Interesting talk. Did you evaluate your approach on any previously established cross-domain few-shot learning benchmarks?

Very nice talk! There are also many works on using computer game engines, such as GTA, to generate labeled datasets, What do you think of this direction in comparison with GAN-based approaches?

Thanks for sharing the awesome works! I have a question about unsupervised pre-training for object detection works (for both ReSim and DETReg). I am wondering whether these two methods are generic to different detector architectures. If not, is it possible to have a unified self-supervised object detection method, which is generic to any kind of detector architecture? What are the fundamental or key properties of this kind of method?

Great Talk! The idea that aligned and uniformity in representation space generalize across tasks & domains seems pretty powerful, especially when we do not know our downstream task a priori. Can we perhaps use these metrics to dynamically explore the multiview/augmentation space (ie. use the metrics to guide something like AutoAugment during the contrastive learning phase)?

Thank you for the talk, this is very interesting work! How were images selected to be annotated for DatasetGAN (i.e., human versus random versus metric)? I would think that this is a very important step to help get the most from each labeled generated image.

Outstanding talk! In the part concerning data privacy, how did you annotate the faces? After you have shown that face obfuscation does not hurt image classification, I would think that the cost of further annotating datasets would be the only reason that a lab/company might not apply this in their work. I would think the cost is very reasonable, but I am not the one making such decisions :). It seems this could be automated with a face detector, but does that just move the problem into the training of the face detector?

Thank you for the wonderful talk! In part 3, the motivation of the idea seems very promising (not all unlabeled data are equal, so we can use a bi-level optimization to weight the unlabeled datapoints). What type of distribution do you observe for the lambda values? I wonder if it would be more close to a bimodal distribution (data is either good or bad), normal distribution, or something else, and what type of insight we could gain from observing this distribution.

Interesting method with impressive results. Two questions: 1) Did you study varying the value of the mixing weight alpha? How does that change results? 2) Did you evaluate your approach in the cross-domain few-shot setting?

Looks like the Convolutional layers are the trick to make the compact transformers faster? What are the effective operators in the transformer to make us choose it over CNNs?

In order to annotate the radar data as suggested by the talk, how can we effectively align the radar data and the pose estimation predictions in temporal space? Since multiple sensors might have their own FPS, what would be the method to align them to make sure the proposed method works?

Excellent works. Could you provide explanations that why the improvements relative to the baseline of ViViT on Kinetics 600 and SomethingSomething V2 are smaller than that on the other datasets? How can we define the proper dataset size to successfully train a model LIke VIVIT?

Thank you for the excellent talks. I have two questions. For ReSim, did you study comparing spatial regions that are not in the same spatial location, by some small perturbations? Because I wonder if enforcing feature consistency over windows that have some degree of jitter may improve performance? For On-Target Adaptation, on the target data you start with a scratch model before contrastive learning -- did you study the impact of starting with various pre-trained models here?

This is an awesome application of few-shot learning with impressive results. I wonder if the approach can generalize across different subject areas? Have you studied this in any other area besides programming? (i.e. perhaps mathematics)?

Fantastic talk, and very interesting insights. Question: can you speculate as to the implications of these findings in the situations where one may be transferring very large models trained on very large datasets, i.e. CLIP, or ViT-G/14 scale?