Seminars

View all Seminars  |  Download ICal for this event

Efficient data analytics using novel algorithms and accelerators

Series: Department Seminar

Speaker: Prof. Tajana Šimunić Rosing, Univ. of California, San Diego

Date/Time: Jan 29 14:00:00

Location: CSA Auditorium, (Room No. 104, Ground Floor)

Abstract:
In today??s world technological advances are continually creating more data than what we can cope with. Traditional computing systems waste a lot of performance moving data from memory and storage for processing. In this talk, I will give an overview of our research which aims to dramatically increase the efficiency of today\'s computers in processing learning tasks. The focus will be on how we utilize emerging computing technologies, both hardware and algorithms, for designing highly-efficient systems. The application areas I will cover include various machine learning, graphs and bioinformatics workloads (genomics, proteomics, metabolomics, drug discovery) . I will first introduce our recent work on accelerating LLMs, multimodal data analytics and transformers using in/near memory processing, followed by the discussion on the efficiency benefits that can be obtained when algorithms are better tuned to the capabilities of hardware. While classical systems rely on 64-bit representation, memory and storage are capable of massively parallel, large vector operations. Hyperdimensional (HD) computing is perfectly suited to this, since its base operations are bitwise and depend on high dimensional random vectors of 1,000s of bits to represent the data. Some of the benefits of HD computing include fast and single pass training, online learning, explainability, robustness to noise, and capability to do symbolic reasoning. I will next provide an overview of my team??s HD computing work with the focus on developments in software and hardware for both big data and edge computing, including: i) novel HDC-based algorithms supporting key cognitive computations in high-dimensional space such as classification, clustering, recommendation systems, secure distributed learning and others, ii) hardware acceleration of workloads using in/near memory and storage, with results of our tapped out chips.

Speaker Bio:
Tajana Šimunić Rosing is a Fratamico Endowed Chair of Computer Science and Engineering and Electrical Engineering, a director of SRC and DARPA funded $56.5M PRISM Center and the System Energy Efficiency Lab at UCSD. She is also an ACM & IEEE Fellow, and was selected as Semiconductor Industry Association’s University Research Award winner for Design in 2022. Her research interests are in energy efficient computing, computer architecture, neuromorphic computing, distributed and embedded systems. She was funded by SRC as a graduate student, and has been involved as a PI and Theme lead in GSRC, MuSyC, TerraSwarm, CRISP, is currently a PI in CoCoSys Center, and leads the PRISM Center. She is also leading a number of DARPA, NSF and SRC funded projects related to HD Computing, SRC funded project acceleration of 3rd generation Fully Homomorphic Encryption, and NSF AI TILOS Research Institute projects on federated learning and AI-based chip design. She was a full time research scientist at HP Labs for 6 years while also leading research efforts at Stanford University. She finished her PhD in EE at Stanford, concurrently with finishing her Masters in Engineering Management. Prior to pursuing the PhD, she worked as a senior design engineer at Intel for 4 years.

Host Faculty: R Govindarajan