Content deleted Content added
→AlphaTensor: fixed grammar/phrasing |
→AlphaTensor: grammar fix |
||
Line 216:
=== AlphaTensor ===
In 2022, [[DeepMind]] introduced AlphaTensor, a [[neural network]] that used a single-player game analogy to invent thousands of matrix multiplication algorithms, including some previously discovered by humans and some that were not.<ref>{{Cite web |title=Discovering novel algorithms with AlphaTensor |url=https://www.deepmind.com/blog/discovering-novel-algorithms-with-alphatensor |access-date=2022-11-01 |website=www.deepmind.com |date=5 October 2022 |language=en}}</ref> Operations were restricted to the non-commutative ground field (normal arithmetic) and [[GF(2)|finite field <math>\mathbb Z/2\mathbb Z</math>]] (mod 2 arithmetic). The best "practical" (explicit low-rank decomposition of a matrix multiplication tensor) algorithm found ran in O(n<sup>2.778</sup>).<ref name="alphatensor">{{Cite journal |last1=Fawzi |first1=Alhussein |last2=Balog |first2=Matej |last3=Huang |first3=Aja |last4=Hubert |first4=Thomas |last5=Romera-Paredes |first5=Bernardino |last6=Barekatain |first6=Mohammadamin |last7=Novikov |first7=Alexander |last8=R. Ruiz |first8=Francisco J. |last9=Schrittwieser |first9=Julian |last10=Swirszcz |first10=Grzegorz |last11=Silver |first11=David |last12=Hassabis |first12=Demis |last13=Kohli |first13=Pushmeet |date=October 2022 |title=Discovering faster matrix multiplication algorithms with reinforcement learning |journal=Nature |volume=610 |issue=7930 |pages=47–53 |doi=10.1038/s41586-022-05172-4 |pmid=36198780 |pmc=9534758 |bibcode=2022Natur.610...47F |issn=1476-4687}}</ref> Finding low-rank decompositions of such tensors (and beyond) is NP-hard; optimal multiplication even for 3x3 matrices [[Computational complexity of matrix multiplication#Minimizing number of multiplications|remains unknown]], even in commutative field.<ref name="alphatensor"/> On 4x4 matrices, AlphaTensor unexpectedly discovered a solution with 47 multiplication steps, an improvement over the 49 required with Strassen’s algorithm of 1969, albeit restricted to mod 2 arithmetic. Similarly, AlphaTensor solved 5x5 matrices with 96 rather than Strassen's 98 steps. Based on the surprising discovery that such improvements exist, other researchers were quickly able to find a similar independent 4x4 algorithm, and separately tweaked Deepmind's 96-step 5x5 algorithm down to 95 steps in mod 2 arithmetic and to 97<ref>{{Cite arXiv |last1=Kauers |first1=Manuel |last2=Moosbauer |first2=Jakob |date=2022-12-02 |title=Flip Graphs for Matrix Multiplication |class=cs.SC |eprint=2212.01175 }}</ref> in normal arithmetic.<ref>{{cite news |last1=Brubaker |first1=Ben |title=AI Reveals New Possibilities in Matrix Multiplication |url=https://www.quantamagazine.org/ai-reveals-new-possibilities-in-matrix-multiplication-20221123/ |access-date=26 November 2022 |work=Quanta Magazine |date=23 November 2022 |language=en}}</ref> Some algorithms were completely new: for example, (4, 5, 5) was improved to 76 steps from a baseline of 80 in both normal and mod 2 arithmetic.
==Parallel and distributed algorithms==
| |||