MSAIL

We at the MichiGAN Student Artificial Intelligence Lab are devoted to reading papers and inviting Yann LeCun to video chat.

[MSAIL @MsailYann]

@boredyannlecun Gradient descent
On parametrized function
Space... Can Learn to Tweet
--- auto haiku generator
#feelthelearn

[Kaggle @kaggle]

Are capsules the future of computer vision? Check out this CapsNet implementation in Kernels benchmarked on MNIST: https://t.co/delejJGmR7

[MSAIL @MsailYann]

@NvidiaAI On 7th day, we rested while the AIs eliminated sickness, poverty, war, and hate. And we saw that it was good. (Or not... @elonmusk) final/

[MSAIL @MsailYann]

@NvidiaAI On 6th day, we programmatically generated architectures.
https://t.co/fxl8rqETGP But: how to learn from uncurated experience? 6/

[MSAIL @MsailYann]

@NvidiaAI On 5th day, we parametrized a subspace of optimizers to learn how to learn. https://t.co/wMptjeim9W But: what architecture to use? 5/

[MSAIL @MsailYann]

@NvidiaAI On 4th day, we developed differentiable GANs, thus learning loss functions. https://t.co/eHWtINpKHc But: how to optimize? 4/

[MSAIL @MsailYann]

@NvidiaAI On 3rd day, we went deep to learn hierarchical features to minimize loss. https://t.co/0gMyxCzQlf But: what loss to use? 3/

[MSAIL @MsailYann]

On 2nd day, we built nets (and machines @NvidiaAI) to combine features automatically https://t.co/wyOd1qbOIJ But: what features to use? 2/

[MSAIL @MsailYann]

On 1st day, since similar ins have similar outs, we wrote rules to instruct machines. https://t.co/d5I6eZiy9d … But: what rules to use? 1/

[MSAIL @MsailYann]

In the beginning, humans hand-classified data. https://t.co/QhPMMCMRFc 0/

[MSAIL @MsailYann]

Stacked Meta-Gaussian Hidden Markov Sparse Autoencoders improve Activity Net Captioning AP by 0.2% #feelthelearn (2/2)

[hardmaru @hardmaru]

Evolutionary Architecture Search, from DeepMind. They represent architectures as genotypes and used GA/random search https://t.co/sjKcZd10j6

[Kyunghyun Cho @kchonyc]

i personally think this may the most (if not only) exciting and successful story of using adversarial learning... https://t.co/cHPHktjrow

[Yann LeCun @ylecun]

Article in TechCrunch about an awesome work from the FAIR-Tel Aviv team about producing avatars that look like a... https://t.co/UooQQWYEIC

[MSAIL @MsailYann]

Psyched for next 3 talks!
oct-31: Memristors
nov-07: Subspace Clustering
nov-14: Invertibility of CNNs
Join us on Tuesdays 6pm in BBB 3725!

[Bored Yann LeCun @boredyannlecun]

Salaries in AI growing fast, but salaries per parameter growing faster, thanks to more convolutions! #feelthelearn https://t.co/ALpakWJPmW

[MSAIL @MsailYann]

@fermatslibrary @balazskegl Apologies: we were wrong and @fermatslibrary was right. As @arunkris95 pointed out, the claim was not as strong as we thought it was.

[MSAIL @MsailYann]

Meet Dr. Jenna Wiens, MSAIL's co-advisor! #TR35 https://t.co/vW6LO7MOd2

[MIT Technology Review @techreview]

This information could help treat patients earlier or prevent C. difficile infections altogether. #TR35 https://t.co/14ulzW982v

[MSAIL @MsailYann]

@fermatslibrary @balazskegl This is incorrect. You conclude that m_x is e^x. However, asymptotically, m should be linear. Mistake is in casework with x vs x-mu < 1.

[Ian Goodfellow @goodfellow_ian]

Impressive high-res GAN results! https://t.co/iwRC1UjkwN (Maybe jump to 2:00 or 4:30)