A Tensorflow to Real-Time Machine Learning (RTML) Compiler

reportActive / Technical Report | Accesssion Number: AD1206486 | Open PDF

Abstract:

This project developed techniques for compiling deep learning models to silicon hardware with a computation mapping and schedule. The key results are (i) a novel technique to effect mapping and scheduling of deep learning model computations on hardware, and (ii) a proof-of-concept energy-efficient hardware implementation capable of both training and inference tasks.

Author(s):
Miesko, Lis
Author Organization(s):
The University of British Columbia
Funding Organization(s):
Defense Advanced Research Projects Agency (DARPA/MTO), Arlington, , VA ; AIR FORCE RESEARCH LAB ABERDEEN PROVING GROUND MD, ABERDEEN PROVING GROUND , MD ; AIR FORCE RESEARCH LAB ABERDEEN PROVING GROUND MD, ABERDEEN PROVING GROUND , MD
Document Type:
Technical Report/Final Report
Publication Date:
2023 Jul 25
Pagination:
36

Security Markings

DOCUMENT & CONTEXTUAL SUMMARY

Distribution Code:
A - Approved For Public Release
Distribution Statement: Public Release

RECORD

Collection: TRECMS
Subject Terms
Modernization Areas:
AI and Machine Learning
Communities of Interest:
Energy and Power Technologies
Descriptor(s):
neural networks, information systems, computers, air force research laboratories, artificial intelligence software, computer languages, computer vision, machine learning, air force, computer programming, information processing, trees (data structures), computer architecture, convolutional neural networks, deep learning, image recognition, measurement
Keyword(s):
machine learning integrated circuit, machine learning algorithms, real-time machine learning, neuroscience-inspired architectures, non-neural ML architectures and generative adversarial learning technique, DEEP LEARNING, ARTIFICIAL INTELLIGENCE, COMPUTER ARCHITECTURE, COMPUTER VISION, STATISTICAL INFERENCE
Subject Categories:
Behavioral and Social Sciences; Mathematical and Computer Sciences; Mathematical and Computer Sciences
Creation Date:
2023 Jul 25
Update Date:
2023 Sep 18