2019’s Top Machine Learning Articles

The best written and video content from the year in machine learning

In this piece, we’ll take a closer look at some of the top machine learning articles in 2019, as ranked by MyBridge.

Facebook, Carnegie Mellon build first AI that beats pros in 6-player poker

According to the article:

Pluribus was tested against professional poker players. This included two winners of the World Series of Poker Event. Pluribus uses self-play to teach itself how to win. This happens without any examples or guidance. The self-play version used in Pluribus is an improved version of the iterative Monte Carlo CFR (MCCFR) algorithm.

Weight Agnostic Neural Networks

The article proposes a search method for neural network architectures that can perform a task without any explicit weight training. These networks are evaluated by populating the connections with single shared weight parameters sampled from a uniform random distribution. The method is able to find minimal neural network architectures that can perform reinforcement learning tasks without weight sharing.

Neural Point-Based Graphics

The authors present a point-based approach for modeling complex scenes. It uses a raw point cloud as the geometric representation of a scene. It then augments each point with a learnable neural descriptor that encodes local geometry and appearance.

According to the paper:

How evolutionary selection can train more capable self-driving cars

Inspired by Darwin’s insights into the evolution, Waymo and DeepMind engineers embark on a research project aimed at making the training of neural networks more effective and efficient. They devise a way to automatically determine good hyperparameter schedules based on evolutionary competition — Population Based Training(PBT).

The networks are evaluated and compete against each other for survival in an evolutionary fashion. A poor performing network is replaced by a better performing network — a progeny. The progeny is a copy of the better performing member of the population.

Introducing Google Research Football: A Novel Reinforcement Learning Environment

Google Research Football is a Reinforcement Learning Environment where agents aim to master football.

It provides a physics-based 3D football simulation where agents can control either one or all football players on their team. They learn how to pass between them, and are able to overcome their opponent’s defense in order to score goals.

The Football Environment provides:

  • A highly-optimized game engine
  • A demanding set of research problems called Football Benchmarks
  • The Football Academy, a set of progressively harder RL scenarios

The beta version can be found on GitHub.

Here’s its paper:

Building Dab and T-Pose Controlled Lights with Python

This article walks you through building a neural network that can recognize dance moves via a webcam. The author uses OpenPose to detect body positions.

The code implementation is available on Github.

Gaining Insights in a Simulated Marketplace with Machine Learning at Uber

The Uber Marketplace Simulation team built a simulation platform that hosts a simulated environment that has riders and driver-partners. The environment mimics scenarios in the real world. It uses an agent-based discrete event simulator. The platform enables Uber Marketplace engineers and data scientists to rapidly prototype and tests new features in a risk-free environment.

Robustness beyond Security: Computer Vision Applications

In this article, the authors discuss the following computer vision paper:

Their hypothesis is that the basic classification framework can be used to tackle challenging image synthesis tasks. The toolkit developed by the authors uses a single, off-the-shelf classifier for all the tasks. The classifier is trained to be adversarially robust.

The code implementation can be found on GitHub:

Video classification with Keras and Deep Learning

This article will teach you how to do video classification using Keras and Python. The article covers how to use a CNN to do video classification.

AI Language Models & Transformers

The video covers how text generation works with a dive into language models and transformers.

Few-Shot Adversarial Learning of Realistic Neural Talking Head Models

This is a video based on this paper:

The paper proposes a model that creates personalized photorealistic talking head models. Its aim is to synthesize photorealistic personalized head images given a set of face landmarks.

Capture the Flag: the emergence of complex cooperative agents

The article covers reinforcement learning in gaming. The authors’ latest paper shows how human-level performance can be achieved in Quake III Arena Capture the Flag.

The authors used a tournament-style evaluation to illustrate that an agent can achieve human-level performance in a three-dimensional multi-player, first-person video game. Only pixels and game points scored are used as input. They used a two-tier optimization process. In this process, a population of independent RL agents are trained concurrently from thousands of parallel matches on randomly-generated environments. Every agent learns its own internal reward signal and rich representation of the world.

End-to-End Deep Reinforcement Learning without Reward Engineering

The article talks about an end-to-end method that allows robots to learn from images that illustrate the successful execution of a certain task. Manual reward engineering is not applied.

The code implementation can be found on GitHub:

Moving Camera, Moving People: A Deep Learning Approach to Depth Prediction

This article tackles depth prediction of moving people. This is done by applying a deep learning-based approach that generates depth maps from videos where people are moving freely. The model learns the human pose and hence doesn’t rely on direct 3D triangulation.

CNN’s, Part 1: An Introduction to Convolutional Neural Networks

This is an introduction to convolution neural networks (CNNs), including a look at how to build one in Python from scratch using NumPy.

Here’s the link to the code implementation:

Advancing self-supervision, CV, NLP to keep our platforms safe

The article covers how Facebook is improving the accuracy and efficiency of its content understanding systems as well as how it’s able to achieve more with less supervised learning.

The main items highlighted from the article are:

  • The self-supervised future of content understanding
  • Record-setting accuracy using hashtags for video understanding
  • Advancing the state-of-the-art in photo and video understanding
  • Using multilingual sentence embeddings to tackle harmful content

Machine Learning Zero to Hero (Google I/O’19)

This is a talk aimed at coders who don’t know machine learning. Some of the concepts covered in the talk are loss functions, optimizers, training loops, and eager mode.

Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask

The authors provide a summary of The Lottery Ticket Hypothesis paper as well as talk about their own paper. Their paper is based on the original Lottery Ticket Hypothesis paper. They also explain why the models in this original paper perform the way they do.


Looking back at 2019, we can certainly see there have been some incredible contributions by authors, research teams, and more in terms of high-quality blog posts, published papers, and explainer videos.



Comments 0 Responses

Leave a Reply

Your email address will not be published. Required fields are marked *