โŒ

Normal view

There are new articles available, click to refresh the page.
Yesterday โ€” 25 June 2024Main stream

Researchers upend AI status quo by eliminating matrix multiplication in LLMs

25 June 2024 at 18:27
Illustration of a brain inside of a light bulb.

Enlarge / Illustration of a brain inside of a light bulb. (credit: Getty Images)

Researchers claim to have developed a new way to run AI language models more efficiently by eliminating matrix multiplication from the process. This fundamentally redesigns neural network operations that are currently accelerated by GPU chips. The findings, detailed in a recent preprint paper from researchers at the University of California Santa Cruz, UC Davis, LuxiTech, and Soochow University, could have deep implications for the environmental impact and operational costs of AI systems.

Matrix multiplication (often abbreviated to "MatMul") is at the center of most neural network computational tasks today, and GPUs are particularly good at executing the math quickly because they can perform large numbers of multiplication operations in parallel. That ability momentarily made Nvidia the most valuable company in the world last week; the company currently holds an estimated 98 percent market share for data center GPUs, which are commonly used to power AI systems like ChatGPT and Google Gemini.

In the new paper, titled "Scalable MatMul-free Language Modeling," the researchers describe creating a custom 2.7 billion parameter model without using MatMul that features similar performance to conventional large language models (LLMs). They also demonstrate running a 1.3 billion parameter model at 23.8 tokens per second on a GPU that was accelerated by a custom-programmed FPGA chip that uses about 13 watts of power (not counting the GPU's power draw). The implication is that a more efficient FPGA "paves the way for the development of more efficient and hardware-friendly architectures," they write.

Read 13 remaining paragraphs | Comments

Before yesterdayMain stream

map โ‰ข territory

By: zamboni
24 June 2024 at 15:02
Welcome to the Principia Mathematica Maps and Table Site (PM-MATS). The goal of this project is to make clear structural connections between different parts of Principia and to make analyzable data about the theorems, definitions, and primitive postulates in its text.

We do this by providing three digital tools: A map of Principia that allows you to see the whole book. 9,944 mini-maps (one for every starred number in Principia) that show you everything used to prove it and everything that it is used to prove (โ‹13.1 for example). A table of Principia that allows users to search for specific starred numbers, sections, chapters, and more, and also allows exportation of search results to JSON or CSV files. Via Trivium

Mathematician Reveals 'Equals' Has More Than One Meaning In Math

By: BeauHD
21 June 2024 at 03:00
"It turns out that mathematicians actually can't agree on the definition of what makes two things equal, and that could cause some headaches for computer programs that are increasingly being used to check mathematical proofs," writes Clare Watson via ScienceAlert. The issue has prompted British mathematician Kevin Buzzard to re-examine the concept of equality to "challenge various reasonable-sounding slogans about equality." The research has been posted on arXiv. From the report: In familiar usage, the equals sign sets up equations that describe different mathematical objects that represent the same value or meaning, something which can be proven with a few switcharoos and logical transformations from side to side. For example, the integer 2 can describe a pair of objects, as can 1 + 1. But a second definition of equality has been used amongst mathematicians since the late 19th century, when set theory emerged. Set theory has evolved and with it, mathematicians' definition of equality has expanded too. A set like {1, 2, 3} can be considered 'equal' to a set like {a, b, c} because of an implicit understanding called canonical isomorphism, which compares similarities between the structures of groups. "These sets match up with each other in a completely natural way and mathematicians realised it would be really convenient if we just call those equal as well," Buzzard told New Scientist's Alex Wilkins. However, taking canonical isomorphism to mean equality is now causing "some real trouble," Buzzard writes, for mathematicians trying to formalize proofs -- including decades-old foundational concepts -- using computers. "None of the [computer] systems that exist so far capture the way that mathematicians such as Grothendieck use the equal symbol," Buzzard told Wilkins, referring to Alexander Grothendieck, a leading mathematician of the 20th century who relied on set theory to describe equality. Some mathematicians think they should just redefine mathematical concepts to formally equate canonical isomorphism with equality. Buzzard disagrees. He thinks the incongruence between mathematicians and machines should prompt math minds to rethink what exactly they mean by mathematical concepts as foundational as equality so computers can understand them. "When one is forced to write down what one actually means and cannot hide behind such ill-defined words," Buzzard writes. "One sometimes finds that one has to do extra work, or even rethink how certain ideas should be presented."

Read more of this story at Slashdot.

if, then

By: HearHere
17 June 2024 at 16:59
George Boole tried to "create a calculus to reduce all logical syllogisms, deductions, and inferences to the manipulation of mathematical symbols, and to cast a precise foundation for the theory of probability. This resulted in his greatest work: An Investigation of the Laws of Thought, [on which are founded the Mathematical Theories of Logic and Probabilities. (Gutenberg, pdf)] a book that laid out the rules of his new symbolic logic and also outlined, in the opening chapter, his grand intention to capture, with mathematics, the language of that ghost that whispers within the tortuous pathways of our minds." [Harper's]

I remember now... These are "quaternions!"

By: Wolfdog
13 June 2024 at 15:08
Imaginary Numbers are Matrices [Japanese with English captions] โ€“ If you would like to have imaginary numbers and quaternions explained in the form of a dialogue between anthropomorphized vocal synthesizers, then Zundamon and Metan are here to oblige you. Zundamon's Theorem is a channel with more of these mathematically enriching conversations.
โŒ
โŒ