The Best of AI: New Articles Published This Month (July 2018)
10 data articles handpicked by the Sicara team, just for you
Welcome to the July edition of our best and favorite articles in AI that were published this month. We are a Paris-based company that does Agile data development. This month, we spotted articles about Reinforcement Learning, Recommendations, Computer Vision and Natural Language Processing. We advise you to have a Python environment ready if you want to follow some tutorials :). Let’s kick off with some good old data humor ’cause not everybody can do modeling for a living:
1 — DeepMind’s New AI Could Be A Better Teammate Than The Average Human At Quake III
Bored of playing with losers who never follow your advice? Let’s step up your game and play with the new DeepMind’s AI at Capture The Flag in Quake III. DeepMind trained this Artificial Intelligence directly from the game’s stream of pixels. The map layout changed at every game. And the AI trained by playing with both other AIs and Humans. In the end, the AI learned close-to-human strategies like “teammate following” or “base camping” from scratch. Researcher lowered the accuracy and increased the time reaction of the AI to prove that its success was not only due to its shooting skills.
Read Capture The Flag — From DeepMind blog
2 — Montezuma’s Revenge By Learning From A Single Human Gameplay
OpenAI trained an agent to achieve a score of 74,500 at Montezuma’s Revenge. The game has not yet been beaten by any Artificial Intelligence though. This blog echoes the idea of “Machine Teaching”. The AI learns based on a limited number of relevant situations coming from the Human gameplay. The blog post does a great job explaining some of the basic principles of Reinforcement Learning. It showcases how a single human-made example alleviates the hurdle of exploring the environment.
Ho! By the way, Sicara has a great introductory article to Deep Q-Learning on Connect 4 working client-side, check it out.
Read Montezuma’s Revenge— From OpenAI blog
3 — How To Build A Search Or Recommendation Engine
This class from Xiangnan He at the National University of Singapore explains the different approaches to building a Search or a Recommendation Engine. This class highlights the common features between search and recommendation. It explains different methods to tackle the issue using Deep Learning or classical algorithms. This class offers a great and complete view of the field.
Read The Slides — From Xiangnan Home
4 — How Tinder Made Its Recommendation Engine More Effective By Using Geolocation
Tinder recommendation engine is based on ElasticSearch. Their user base is huge but only a few users can effectively match because they are located nearby. They solved this issue by relying on what they called Geosharding. Geosharding is a location-based ElasticSearch cluster sharding using Google S2. This is an inspiration in new ways to improve your data engineering solutions or to tackle location-based problems.
Read GeoSharding — From Tinder blog
5 —How Netflix Personalizes Your Whole Home Page
Netflix is well-known for its personalized website. Your landing page ranks the movies but also finds the most relevant artworks for you. Last month the company held their Personalization, Recommendation, and Search (PRS) workshop. Here is a cool “takeaways” post about this event. You might be interested in Multi-Armed Bandits, how to deal with the Survivor Bias or the Tyranny of the Majority. Or you might want to know how to solve real-world recommendation problems while keeping in mind scalability. This blog and this event are filled with good insights.
Read Takeaways From PRS — From Van Der Goes’ Medium
6 — Waymo Is Hitting The Road In Both Real Life And Simulation
Waymo, the former Google self-driving car project aims to create “The World’s most experienced driver”. Waymo’s cars are driving 25,000 miles a day while the company is simulating another 1,500,000 miles. The past months have been harsh for the autonomous vehicle industry with the so-called Winter of self-driving AI. More recently Uber was forced to shutter its self-driving truck project, Otto. In this talk, Waymo CEO John Krafcik explains some of the issues the self-driving industry is facing and how Waymo plans to overcome them. Self-driving is facing a lot of problems from the well-known safety issues to User Experience ones like how and where to pick-up passengers.
Read Self Driving AI Winter — By Russel Brandom
Read Uber shutters its self-driving truck project — By Andrew J. Hawkins
Read Waymo’s autonomous vehicles are driving 25,000 miles every day — From Kirsten Korosec’s Tech Crunch
Watch Krafcik speaking about Waymo— From NGA’s YouTube Channel
7 — Switchable Normalization
Normalization is a great improvement in a lot of Deep Learning system. But like activation functions, there are a lot of them and it’s hard to choose from. Most data scientists choose one normalization above the others by habit or lack of knowledge of the particular situation. Most of the time a normalization will be chosen for all the network rather than each situation. This is by design sub-optimal. Switchable normalization has been introduced to solve this issue. This is a trainable normalization layer that will find the best normalization layer for your network at each point of interest.
Right now switchable norm is switching between three normalizations:
- batch normalization: classic after convolutional layers;
- instance normalization: when having high variance across batches;
- layer normalization: better after recurrent layers;
The reasoning behind this choice is double. First, those are frequently used and known to be effective. Second, some parts of the computation can be shared which permit to use Switchable Norm without adding a huge burden to the neural network.
Read Differentiable Learning-to-Normalize via Switchable Normalization — By Russel Brandom
See Switchable Normalization Implementation with PyTorch— From Switchnorm Github
8 — Glow: Reversible Generative Models
OpenAI introduces Glow, a reversible generative model based on invertible 1x1 convolutions.This approach shows great results in term of quality but needs a lot of layers to produce high-quality images. In the future, OpenAI’s searchers already plan to make Glow more efficient regarding the computing time and the number of parameters. The blog post has an awesome User Interface to play around with the parameters an see Glow’s results.
Read Glow — From OpenAI Blog
9 —Fixing Noisy Images By Training Only On Noisy Images
NVIDIA’s new AI achieves impressive Denoising Results with learning on noisy images.
In this paper, NVIDIA’s team introduces an algorithm which learns how to produce a higher-quality image, without ever “seeing” one.
This is like teaching someone how to clean a room showing him a way to make the room cleaner rather than showing him what a clean room looks like. This article shows how a simple statistical idea can have a great impact on machine learning problems.
Read Cleaning Grainy Images — From NVIDA Blog
10—Learning A New Task In A New Language With Lower Resources
Facebook’s research team has tackled the multilingual multi-task problem: how to cope with training a model for a task and a language you have low resources for. Their answer lies in the use of shared units between task and/or language. You may not be interested in this specific task but the paper is interesting for various reasons. First, it constitutes a good introduction to various units and methods used in Machine Learning and Natural Language Processing in particular: bidirectional LSTM (well-explained here), CRF, Highway Network, Character-level CNN. The paper explains why one would want to use those units and how it relates to task-specific or language specific features. Second, it shows how to think and build a neural network for your needs using your experience.
Read A Multi-lingual Multi-task Architecture for Low-resource Sequence Labeling — From Facebook Research Blog
We hope you’ve enjoyed our list of the best new articles in AI this month. Feel free to suggest additional articles or give us feedback in the comments; we’d love to hear from you! See you next month.
Read the June edition
Read the May edition
Read the April edition
Read the March edition
Read the February edition
Read the original article on Sicara’s blog here.