📝 Editorial 

Gaming has been at the center of the ML renaissance for the last few years. In many ways, the deep learning race was kicked off by AlphaGo performance again Go’s legend Le Sedol. However, most of the advancements in ML for gaming have been specialized in either perfect or imperfect information games but never both at the same time. Models that mastered chess and Go struggle with imperfect games such as Poker. Even models like AlphaZero, which learned to play multiple games simultaneously, were constrained to perfect information environments. The reason for this is based on the intrinsic dynamics of both types of game environments. Perfect information games, like Chess and Go, are a good fit for ML techniques relying on self-play learning and local min-max search of the gamespace, while imperfect information games like Poker rely on game-reasoning techniques. Earlier this week, DeepMind unveiled a new ML method that can change these constraints forever.  

Player of Games (PoG) is the first ML model able to play perfect and imperfect information games at scale. Created by DeepMind, PoG combines self-play learning, search, and game-theoretic reasoning in a single model that can adapt to perfect and imperfect information environments. The model achieved super-human performance in perfect games like Chess and Go as well as imperfect games like Poker or Scotland Yard. Beyond the applications in gaming, PoG is a massive leap towards building ML models that can adapt to real-world domains such as weather forecasting or energy optimization that combine perfect and imperfect information. We should expect PoG to become a seminal algorithm in the next wave of ML gaming research.  

Share

🔺🔻 TheSequence Scope is our Sunday free digest. To receive high-quality educational content about the most relevant concepts, research papers and developments in the ML world every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻

🗓 Next week in TheSequence Edge:

Edge#149: we discuss Model Tracing and Lineage; we explore MLTrace, a reference architecture for observability in ML pipelines; we overview M3, a platform that powers time-series at Uber.    

Edge#150: we deep dive into Microsoft’s SynapseML, a new framework for large-scale ML.

Now, let’s review the most important developments in the AI industry this week

🔎 ML Research

Playing Perfect and Imperfect Games 

DeepMind published a paper detailing Player of Games, an algorithm that can play both perfect and imperfect information games →read more in the original research paper 

Language Models at Scale 

DeepMind published three papers detailing new techniques to scale language models →read more on DeepMind blog

Debugging ML Models 

Amazon Research published a paper proposing a method for debugging ML models →read more on Amazon blog

Improving Transformers for Computer Vision 

Google Research published a paper proposing a technique to optimize transformers for computer vision by using a smaller number of tokens →read more on Google Research blog

🤖 Cool AI Tech Releases

MiniTorch  

The code for MiniTorch, a library for learning the principles of ML in Python, was open-sourced →read more on MiniTorch’s Github

🛠 Real World ML

Testing Firefox 

Mozilla published a blog post detailing the ML architecture they use to test Firefox more efficiently →read more on Mozilla blog 

LinkedIn Analytics Stack 

LinkedIn published a blog post describing the evolution of their analytics stack →read more on LinkedIn blog

Large Language Models for Shopping Recommendations 

Korean search giant Naver is using massive language pretrained models to power shopping recommendations →read more in this article from VentureBeat  

🗯 Useful Tweet

DeepMind @DeepMind

Today at #NeurIPS2021: 🔷@WiMLworkshop & @QueerinAI workshops & socials 🔷 A framework for Verifying Probabilistic Specifications 🔷 Variational Bayesian Optimistic Sampling 🔷Collaborating with Humans without Human Data See the full schedule via dpmd.ai/neurips2021

December 9th 2021

20 Retweets78 Likes

Follow us on Twitter

💸 Money in AI

ML&AI

• Development platform for embedded ML Edge Impulse raised $34 million in a Series B round led by Coatue. Hiring remote.

• Model vulnerabilities detection and prevention startup Robust Intelligence raised a $30 million Series B funding round led by Tiger Global. Hiring in San Francisco/US.

AI-powered

• Voice platform for healthcare professionals Suki raised $55 million in a Series C round of funding led by March Capital. Hiring in Bengaluru/India and Redwood City/US.

• Talent engagement platform Sense raised $50 million in series D funding led by SoftBank. Hiring in Bengaluru/India and San Francisco/US.

• AI video generation platform Synthesia raised $50 million in a Series B round led by Kleiner Perkins. Hiring globally.

• Customer engagement platform MoEngage raised a $30 million Series D financing round led by Steadview Capital. Hiring globally.

• Conversational AI company Posh Technologies raised $27.5 million in Series A funding led by Canapi Ventures. Hiring in Boston/US.

• Cloud analytics application platform NetSpring raised $13 million in a Series A round led by Dell Technologies Capital. Hiring in San Francisco/US.