Machine Learning
Projects
Portable High Performance Inference at the Tactical Edge
In collaboration with the Software Engineering Institute
This project focuses on the development of a portable software
framework for rapidly instantiating high performance ML libraries for
devices with SWAP constriants. Specifically, the framework targets
edge devices, and incorporates inference-specific optimizations such
as aggressive layer fusion.
DARPA TRIAD: CAMP-COTS Co-designed Array and ML processing on COTS architectures
Co-PIs: Yuejie Chi, James Hoe, Swarun Kumar
This project focuses on an exploration and demonstration of
the technology advancements that are requried to bring about
intelligent signal processing using commercial-of-the-shelf (COTS)
computational devices. The development of the proto-type CAMP-COT
system will demonstrate the software-hardware stack that is needed
to facilitate the interactions between the ML community, signal
processing community, and the high performance community to meet
the objective of TRIAD.
Analytical Models for Efficient Data Orchestration in DL Workloads
In collaboration with Facebook
News
This project focuses on the design and use of analytical models
to inform the design of high performance deep learning network
implementations. As the deep learning networks can be composed
and configured in different manner, it is essential that there is
a systematic approach to orchestrate data movement through the
different layers in a deep learning network. This approach will
allow facilitate the portability of the deep learning network
across different configurations and computing platforms.
Speeding up Computer Vision for Conveyer Line Processing
In collaboration with CMKL University (Thailand)
Practical deep learning based classification must be accurate, and must
be performant to meet operational needs.
This project focuses on the design and optimization of deep
learning models that are used to aid the identification of
recyclable bottles. These bottles are collected from different
sources, and ML-driven computer vision is used to classify the
conditions of the collected bottles. Another objective of this
project is to speed up the classification process such that the
classification can be performed at conveyer line speed.
Publications
Fusing Non Element-wise Layers in DNNs (Extended Abstract & Poster)
Upasana Sridhar, Tze Meng Low, Martin Schatz
IEEE High Performance Extreme Computing Conference (HPEC)
2021
High Performance Zero-memory Overhead Direct Convolutions
Jiyuan Zhang, Franz Franchetti, Tze Meng Low
International Conference on Machine Learning
2018
