Xarrada ELLIS 'Fast approximate matrix multiplication – theory and practice in AI hardware'
Visites: 2139
Comentaris a la notícia: 0
La xarrada serà impartida pel Prof. Nir Ailon, el 7 d'abril a les 11.30 en la Sala d'Actes de la Politècnica IV.
ELLIS DISTINGUISHED SPEAKER TALK
Títol: Fast approximate matrix multiplication – theory and practice in AI hardware
Ponent: Prof. Nir Ailon
Lloc: Saló Actes de la Politecnica IV
Data: 07-04-2022 d'11.30 a 13.00
Abstract: Multiplying large, dense matrices is a key ingredient of deep networks, and state of the art architectures may spend more than 99% of time and energy consumption on matrix multiplication. Although fast matrix multiplication algorithms in time o(n^3) are known in the theoretical literature, they are impractical, and not used in the AI hardware and software ecosystem today.
In this talk I will survey algorithms and heuristics for faster approximate matrix multiplication, and show the possible impact, as well as limitations, of such approaches on the future of deep learning. The results I will present will cover both academic research, as well as my more recent experience with cutting edge AI industry at deci.ai, where I have discovered a huge gap between the academic research and practical AI.
Short Bio: Nir Ailon is a Computer Science professor at Technion in Haifa Israel. He completed his PhD in 2006 at Princeton University, and then continued as a postdoc at the Institute for Advanced Study in Princeton. He has served as faculty at Technion since 2011, and also spent time at Google Research, Yahoo! Research, as well as at other industrial research labs. He is a recipient of an ERC Consolidator Grant. His research spans theory of algorithms, mathematical foundations of big data and machine learning. He is currently involved in research on AI acceleration at deci.ai
Comentaris a la notícia 
Nota:
Aquesta és l'opinió dels usuaris, no de l'EPS.
No està permès abocar comentaris contraris a les lleis espanyoles o injuriantes.
Reservat el dret a eliminar els comentaris que considerem fóra de tema.
Els comentaris s'ordenaren dels més recents als més antics.
