Final GANtasy

Grad student at UIUC, aimlessly wondering what project to move on to next. And no, I haven't used GANs.

[Tanmay Gupta @tanmay2099]

Feels a bit weird not being a graduate student anymore, but that was Day 1 as a Research Scientist at Team PRIOR @allen_ai 👨‍🔬

Looking forward to working with amazing colleagues @anikembhavi @RoozbehMottaghi @LucaWeihs @ehsanik @jiasenlu and many others!

[Final GANtasy @lizhitwo]

, which is the point of scientific discussions — learn from each other. On Twitter, though, the risk of side effects is high.

[Final GANtasy @lizhitwo]

My lesson from many many Twitter debates: Twitter cannot be used for discussions. Period.

Once you have many followers, even if people discussing with you assume good intent, others may not. And all of your words WILL be taken out of context, accidentally or intentionally. So..

[Yann LeCun @ylecun]

@anmarasovic @timnitGebru @soumithchintala No. I don't need any applause.
Assumption of good intent would be nice though.

How can science and understanding progress if no one is allowed to explain their ideas?
As an (African) friend recently told me, "the engine of science is the celebration of constructive disagreement"

[Final GANtasy @lizhitwo]

So how can we balance our roles as (1) people (2) brands (3) researchers (4) educators (5) public figures (6) citizens all at the same time? Under the nose of millions of users, no less.

I’m sure most of this discussion is fine in private and would lead to mutual understanding,

[Andrej Karpathy @karpathy]

Important to be aware of & mitigate model calibration in real applications where thresholded outputs lead to diff behaviors. Good CVPR2020 paper on the topic: https://t.co/DTVWZLzsL9 TLDR: use an ensemble of T-scaled models, distill if you can't afford the added inference latency

[Geoffrey Hinton @geoffreyhinton]

Organizers of data science and machine learning conferences (NeurIPS, ICML, AISTATS, ICLR, UAI, ...): Allow remote paper & poster presentations at conferences - Sign the Petition! https://t.co/gerNJ7lxML via @Change

[Final GANtasy @lizhitwo]

Check out our newest paper on synthesizing the training images from just a classifier!

[Arun Mallya @arunmallya]

Do Deep Networks dream of (Electric) Sheep?
Yes, they do!
Introducing #DeepInvert (https://t.co/SQ2voDVid4). It allows the synthesis of previously used training data, given just a trained model.
*No new model training, no training metadata required.*

[Arun Mallya @arunmallya]

Do Deep Networks dream of (Electric) Sheep?
Yes, they do!
Introducing #DeepInvert (https://t.co/SQ2voDVid4). It allows the synthesis of previously used training data, given just a trained model.
*No new model training, no training metadata required.*

[Arun Mallya @arunmallya]

Here’s what the PowerPoint automatic caption generator has to say – “A herd of sheep standing on top of a rocky mountain”!

[Arun Mallya @arunmallya]

Based on @zzznah’s #DeepDream, our method performs intermediate feature regularization using stored batchnorm statistics.We are able to generate all previously seen training classes with unmatched realism and fidelity.

[hardmaru @hardmaru]

Using gradient descent for everything https://t.co/okUR4RKTiQ

[Final GANtasy @lizhitwo]

Since the “transformer” name in the Transformer Network doesn’t mean transforming or anything, someone should totally write a Avenger Network paper

[Final GANtasy @lizhitwo]

🕯️

[Unnat Jain @unnatjain2010]

Our CVPR 2019 (oral) paper is now available on arxiv!
Project page: https://t.co/kRqyTFLcn4
Paper: https://t.co/j2mrcGtG5S
Joint work with @allen_ai @IllinoisCS @xnor_ai @uwcse