Pitfalls of Batch Norm in TensorFlow and Sanity Checks for Training Networks

Some more practical advices on #tensorflow training with source code and reference links to look up.

https://medium.com/@theshank/pitfalls-of-batch-norm-in-tensorflow-and-sanity-checks-for-training-networks-e86c207548c8

#beginner #novice #dl #tutorial

Introduction to Python Ensembles

Ensemble is an approach for mixing algorithms to obtain strong sides of all the approaches.

https://www.kdnuggets.com/2018/02/introduction-python-ensembles.html

#ensemble #stacking #tutorial #beginnig #novice

Adversarial attack — type of input or a mask applied to the input of the machine learning model to make it wrong. It is a way to cheat with the output, to ‘fool’ the algorithm.

«Attacking Machine Learning with Adversarial Examples» at Open AI blog covers the basics and provides some examples.

Open AI blog article: https://blog.openai.com/adversarial-example-research/

#adversarialattack #openai #novice #beginner

New attack on neural networks can alter the purpose of the neural network.

A surprising adversarial attack, whereby a perturbation to all input images can "reprogram" a poorly-defended neural network to change its task entirely. e.g. turn an ImageNet classifier into a network that counts squares.

Arxiv: https://arxiv.org/pdf/1806.11146.pdf

#Goodfellow #gbrain #adversarialattack

Most common pitfalls, you can encounter when training neural network.

http://telegra.ph/Most-common-neural-network-mistakes-07-01

#beginner #novice #dl #tutorial

Deep Mind announced that its agent beated human performance in Quake III CTF (Capture The Flag)

https://deepmind.com/blog/capture-the-flag/

#rl #quake3 #deepmind

ModaNet: A Large-Scale Street Fashion Dataset with Polygon Annotations

Latest segmentation and detection approaches (DeepLabV3+, FasterRCNN) applied to street fashion images. Arxiv paper contains information about both: net and dataset.

Arxiv link: https://arxiv.org/abs/1807.01394
Paperdoll dataset: http://vision.is.tohoku.ac.jp/~kyamagu/research/paperdoll/

#segmentation #dataset #fashion #sv

Hey, our fellow colleagues at OpenDataScience community are labeling a meme dataset. You can help them with the markup just by viewing memes in this bot: @MemezoidBot

#DataSet #labeling

Neural scene representation and rendering

In June #DeepMind introduced Generative Query Network (#GQN) framework within which machines learn to perceive their surroundings by training only on data obtained by themselves as they move around scenes.

Link: https://deepmind.com/blog/neural-scene-representation-and-rendering/

#DeepMind new release: Neural Processes (#NPs) that generalise #GQN ’s training regime to other few-shot prediction tasks such as regression and classification

Arxiv 1: https://arxiv.org/abs/1807.01622
Arxiv 2: https://arxiv.org/abs/1807.01613

#ICML2018

How #Netflix used data science to make scripts for the shows:

Video: https://www.youtube.com/watch?v=qXo9jTxfqJ8&feature=youtu.be
Github: http://netflix.github.io

Really great video, showing practical approach, with some focus on human interactions and integrating data insights into the product.

#youtube #netflixresearch

Udacity has published a github repo for the Deep Reinforcement Learning Nanodegree program

Repo: https://github.com/udacity/deep-reinforcement-learning
Nanodegree: https://www.udacity.com/course/deep-reinforcement-learning-nanodegree--nd893

#dl #udacity #mooc #course #github #rl

Deep Learning for Matching in Search and Recommendation

PDF: http://www.comp.nus.edu.sg/~xiangnan/sigir18-deep.pdf

#sigir2018 #Tutorial

Glow by #OpenAI: Better Reversible Generative Models

Project link: https://blog.openai.com/glow/
Video: https://d4mucfpksywv.cloudfront.net/research-covers/glow/videos/both_loop_new.mp4

Dynamic few-shot visual learning without forgetting by Gidaris and Komodakis

The authors share even configs and learning rate schedules for experiments in paper.

Github: https://github.com/gidariss/FewShotWithoutForgetting

#cvpr2018

SwitchNorm: add BatchNorm + InstanceNorm + GroupNorm with a learnable blend at each layer.

Paper about optimal normalization in neural nets continues. Plots + code

Arxiv: https://arxiv.org/abs/1806.10779

#dl #normalization

And it would be a crime to leave #ICML2018 out.

It is held in Stockholm now, papers are available at: http://proceedings.mlr.press/v80/

#conference

Toward theoretical understanding of deep learning

http://unsupervised.cs.princeton.edu/deeplearningtutorial.html

#ICML2018 #tutorial #dl