Exploring Fault Tolerant Topologies of Neural Networks in New Technologies of All Programmable Syste
Abstract
This research project proposes the investigation of the vulnerability of the new all-programmable logic devices (APSoC) fabricated i,n Mosfet and Finfet technologies in-applications related to artificial intelligence (AI) implemented as neural networks operating in,-harsh environments. Today, state-of-the-art programmable logic devices, known as Field-Programmable Gate Array (FPGA) devices, incl,ude not only the programmable logic fabric-but also hard-core parts, such as general-purpose processors, dedicated processing blocks,,-interfaces to various peripherals, on-chip bus structures, and analog blocks. These new-devices are commonly called All Programmab,le System-on-Chip (APSoC) devices. APSoC-devices have a large amount of computational resources available, which make them very-attr,active to embedded neural networks for pattern and image recognition for unmanned aerial-vehicles, military, earth observation satel,lites and aerospace applications in general. Tasks-based on image detection and classification, complex math calculations, sensor fu,sion and-route prediction are some examples of real-time applications that can run in those devices.-However, APSoC can be very susc,eptible to radiation-induced faults, and they are very-complex chips to test due to the amount of different architectures and resour,ces. APSoC-needs to be qualified under faults not only by static testing but also by dynamic testing-considering the characteristics, of the AI application. There are many potential neural network-topologies and it is challenging to find the best neural topology wi,th the most efficient fault-tolerant technique matching the requirements of a specific task with best accuracy and-resource usage in, terms of time, power, area and reliability
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 08, 2022
- Source ID
- N629092212014
Entities
People
- Fernanda Kastensmidt
Organizations
- Federal University of Rio Grande do Sul
- Office of Naval Research
- United States Navy