🧮 DeepMind’s AlphaTensor can Discover New Math Algorithms
Weekly news digest curated by the industry insiders
📝 Editorial
Algorithms have been the cornerstone of mathematics since ancient times. The world Algoritmi was used by Persian mathematician Muhammad ibn Musa al-Khwarizmi to describe some of his work with linear and quadratic equations. Since then, the algorithms have been used to describe sequences of operations that produce a target output. The process of discovering new algorithms in mathematics requires not only knowledge but reasoning and intuition, which have long been considered exclusive to the human mind. Despite the advent of the computer era, that hasn’t changed much, which provides an idea of how complex the process is.
So can AI be used to discover new algorithms?
Last week, DeepMind announced a giant leap forward in this space with the unveiling of AlphaTensor, a new model that discovers new algorithms in fundamental areas such as matrix multiplications.
AlphaTensor is the evolution of AlphaZero, DeepMind’s agent that achieved superhuman performance on board games, but applied to mathematical problems. Matrix multiplication seems like the perfect area to focus on, given that it is one of the foundations of machine learning. In early algebra classes, we learn a simple algorithm for matrix multiplication that infers a linear operation from the matrix structure. For centuries, mathematicians thought that algorithm to be the most efficient method for matrix multiplications until German mathematician Volker Strassen showed a more optimal approach in 1969. Since then, the math community has been trying to discover more efficient matrix multiplication methods for large matrices, common in domains such as computer vision and speech analysis. To train AlphaTensor, DeepMind redefined a matrix multiplication problem as a single-player game that scores the algorithm’s efficiency. The number of potential solutions to a given board composition is larger than the number of atoms in the universe. Building on the AlphaZero principles, AlphaTensor uses a reinforcement learning method that mastered the game by just playing it. The result was the discovery of not one but many matrix multiplication methods that are far more efficient than the established ones.
AlphaTensor represents a major leap forward in algorithm discovery. The idea of modeling math problems as a game and having an ML agent discover new algorithms is certainly novel and can become the foundation for advancing many fields in mathematics.
🔺🔻TheSequence Scope – our Sunday edition with the industry’s development overview – is free. To receive high-quality content about the most relevant developments in the ML world every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻
🗓 Next week in TheSequence Edge:
Edge#233: we explain DALL-E 2; discuss the DALL-E 2 paper; explore DALL-E Mini (Now Craiyon), the most popular DALL-E implementation in the market.
Edge#234: we deep dive into Meta AI’s Make-A-Video: the new super model that generates videos from textual inputs.
📌 ML:Integrity conference / Oct 19
Join us at ML:Integrity on Oct 19th! It is a free, virtual conference dedicated to advancing machine learning integrity.
Learn from those building strategies and engineering paradigms to eliminate ML model failure.
Attend over a dozen sessions on ML fairness, security, scale, regulation, open source, and more.
Participate in a forum that offers education, inspiration, and progress toward furthering the cause of ML integrity.
Now, let’s review the most important developments in the AI industry this week
🔎 ML Research
AlphaTensor
DeepMind published a research paper detailing AlphaTensor, the first extension of AlphaZero to mathematics that is able to discover novel algorithms →read more
Imagen Video
Google followed Meta by publishing a paper detailing Imagen Video, its own text-to-video generative model →read more
FILM
Google Research published a paper introducing FILM, a method for creating slow-motion videos from near duplicate photos →read more
🤖 Cool AI Tech Releases
AITemplate
Meta AI open-sourced AITemplate, an inference framework that provides hardware acceleration for both NVIDIA and AMD GPUs →read more
Domino 5.3
Domino Data Labs announced the release of a new version of its data science platform with new features for MLOps, multi-cloud support and accelerated inference →read more
GraphOS
GraphQL platform Apollo launched GraphOS, a new platform to build “supergraphs” or massive connected data structures that integrate data from different sources →read more
Mintaka
Amazon Science open-sourced Mintaka, a dataset for multilingual question answering →read more
🛠 Real World ML
The State of Conversational AI
Salesforce Research published an insightful blog post detailing the history, current state and future of conversational AI →read more
Meta’s Feed Optimization
Meta AI detailed the ML methods powering the new Show More or Show Less features in the Facebook app →read more
element platform at Walmart
Walmart published details about element, the platform powering their internal ML workflows →read more
GraphQL at LinkedIn
LinkedIn discusses the evolution of their GraphQL infrastructure and some of the best practices implemented in their internal architecture →read more
💸 Money in AI
AI-powered
Supply chain tech startup Altana raised a $100 million Series B investment led by Activate Capital. Hiring in Brooklyn/US and London/UK.
Conversational AI assistant Xembly raised a $15 million Series A funding round led by Norwest Venture Partners. Hiring in Seattle/US.
Supply chain robotics startup Gather AI raised a $10 million Series A round led by Tribeca Venture Partners. Hiring in India and the US.
Acquisitions
Content moderation startup Oterlu was acqui-hired by Reddit for an undisclosed amount.
Telerobotic chessboard creator Square Off was acquired by consumer robotics company Miko for an undisclosed amount.