​​Testing Robustness Against Unforeseen Adversaries

OpenAI developed a method to assess whether a neural network classifier can reliably defend against adversarial attacks not seen during training. The method yields a new metric, #UAR (Unforeseen Attack Robustness), which evaluates the robustness of a single model against an unanticipated attack, and highlights the need to measure performance across a more diverse range of unforeseen attacks.

Link: https://openai.com/blog/testing-robustness/
ArXiV: https://arxiv.org/abs/1908.08016
Code: https://github.com/ddkang/advex-uar

#GAN #Adversarial #OpenAI

OpenGPT-2: We Replicated GPT-2 Because You Can Too

Article about replication of famous #GPT2. This replication project trained a 1.5B parameter «OpenGPT-2» model on OpenWebTextCorpus, a 38GB dataset similar to the original, and showed comparable results to original GPT-2 on various benchmarks.

Link: https://medium.com/@vanya_cohen/opengpt-2-we-replicated-gpt-2-because-you-can-too-45e34e6d36dc
Google colab: https://colab.research.google.com/drive/1esbpDOorf7DQJV8GXWON24c-EQrSKOit
OpenWebCorpus: https://skylion007.github.io/OpenWebTextCorpus/

#NLU #NLP

Open-sourcing hyperparameter autotuning for fastText

Facebook AI researchers are releasing a new feature for the fastText library which provides hyper-parameter autotuning for more efficient text classifiers.

Link: https://ai.facebook.com/blog/fasttext-blog-post-open-source-in-brief/

#FacebookAI #Facebook #FastText #NLU #NLP

​​The infinite gift

is an interesting object where the side of the nth box is 1/√n. As n→+∞, the gift has infinite surface area and length but finite volume!

​​Exploring Weight Agnostic Neural Networks

Exploration of agents that can already perform well in their environment without the need to learn weight parameters.

Link: https://ai.googleblog.com
Code: https://github.com/google/brain-tokyo-workshop/tree/master/WANNRelease

​​Neural net to enhance old or low-quality video to HD (TS -> HD).

It is so surprising that noone had yet released a model for that. People have lots of old video recordings, which will definately benefit from quality enhancement. And we all have to hope movie pirates won’t use it to enhance stolen copies.

Link: https://news.developer.nvidia.com/researchers-at-videogorillas-use-ai-to-remaster-archived-content-to-4k-resolution-and-above/
More demos: https://videogorillas.com/bigfoot/

#SuperResolution #CV #DL

ODS breakfast in Paris! See you this Saturday at 10:30 at Malongo Café, 50 Rue Saint-André des Arts.

http://rescience.github.io/
Tl;dr:
Reproducibility is important. Publishing a paper which results can't be used by any reader is more or less useless. However, while everybody talks about reproducibility, but nobody accepts papers about reproduction of the existing research for publication, let alone the fact of publishing non-reproducible research (not enough details, no open dataset, etc.), which is OK sometimes, but usually is not.
Moreover, what people usually mean when they say "reproducibility" (possibility of repeating the exact experiment described in paper and achieving same results) is "replicability" (possibility of conducting similar experiments with similar results).
This journal aims to be an open access and open source platform to publish replication computational research (which is easier to both replicate and verify).

🚨😭STOP talking bad about different Data SPECIALTIES😭🚨

Data Science is EXCITING

Frequentist Statistics is RELIABLE

Software Engineering is CRUCIAL

Bayesian Statistics

Machine Learning is POWERFUL

​​New fastMRI challenge from #FacebookAI team

Submission deadline: September 19

Announcement link: https://ai.facebook.com/blog/fastmri-challenge/
Competition link: https://fastmri.org/

#Competition #NotOnlyKaggle #Facebook #CV #DL

Nice article on non-official jupyter notebook extensions

Warning: there is a checkbox, saying «disable configuration for nbextensions without explicit compatibility (they may break your notebook environment, but can be useful to show for nbextension development)». So it is better to test the extensions in separate environment.

And correct way to install is extension support is:

pip install jupyter_contrib_nbextensions && jupyter contrib nbextension install --user

Link: https://towardsdatascience.com/setting-up-a-data-science-environment-using-windows-subsystem-for-linux-wsl-c4b390803dd

#jupyter #tipsandtrics

​​And the Bit Goes Down: Revisiting the Quantization of Neural Networks

Researchers at Facebook AI Research found a way to compress neural networks with minimal sacrifice in accuracy.

Works only on fully connected and CNN only for now.

Link: https://arxiv.org/abs/1907.05686

#nn #DL #minimization #compresson

Great collection and reviews for top online machine and deep learning courses

Post covers short reviews and suggested order in which course could be taken, along with the links at great prerequisites.

Link: http://thegrandjanitor.com/2016/08/15/learning-deep-learning-my-top-five-resource/

#DL #ML #MOOC #novice #entrylevel

Posture detector running on #tensorflowjs

Great example of learning project for common good. Detects posture and blurs screen if posture is not right.

Link: https://fix-posture.glitch.me
Design: https://www.figma.com/file/hvYOl9g4oO2UaTpVOqRwcF/AI-experiments-by-me-Good-posture?node-id=0%3A1
Code: https://glitch.com/~fix-posture

#posenet #poseestimation

​​Understanding Transfer Learning for Medical Imaging

ArXiV: https://arxiv.org/abs/1902.07208

#biolearning #dl #transferlearning

ODS breakfast in Paris! See you this Saturday (Tomorrow) at 10:30 at Malongo Café, 50 Rue Saint-André des Arts.

Deep Fake Challenge by Facebook team

#Facebook launches a competition to fight deep fakes. Unfortunately, results of this competition will be obviously used to create better fakes, to the cheers of the people, wishing to watch the Matrix with Bruce Lee or more questionable deep fake applications.

Link: https://ai.facebook.com/blog/deepfake-detection-challenge/

#deepfake #video #cv #dl

​​An Easy Introduction to Generative Adversarial Networks

GANs are main type of neural networks used for data generation.
#intro level post on how #GAN s work

Link: https://towardsdatascience.com/an-easy-introduction-to-generative-adversarial-networks-6f8498dc4bcd

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

How normalization applied to layers helps to reach faster convergence.

ArXiV: https://arxiv.org/abs/1502.03167

#NeuralNetwork #nn #normalization #DL

​​Using Deep Learning to Inform Differential Diagnoses of Skin Diseases

Deep Learning System (DLS) for quicker and cheaper skin diseases detection. DLS showed accuracy across 26 skin conditions on par with U.S. board-certified dermatologists, when presented with identical information about a patient case (images and metadata). This is an amazing example of how technology can help fight notoriously high medical bills in the USA and make top-level care available and more affordable in all other the world.

Link: https://ai.googleblog.com/2019/09/using-deep-learning-to-inform.html?m=1
ArXiV: https://arxiv.org/abs/1909.05382

#Inception4 #Google