NETRIUS Deep Chaos

reportActive / Technical Report | Accesssion Number: AD1153246 | Open PDF

Abstract:

Artificial intelligence, and deep learning in particular, is increasingly becoming a critical component in our national infrastructure. It is imperative that techniques used to train these models be consistent in their predictions. Fooling these systems with small perturbations to the input data is a well-known problem. Identifying and characterizing regions of sensitivity is less understood in the Al domain, but well studied in nonlinear dynamics mathematics. We seek to investigate whether deep neural networks are chaotic, and, if so, can they be stabilized?

Author(s):
Eslami, Mohammed
Author Organization(s):
Netrias, LLC
Funding Organization(s):
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH ARLINGTON VA, ARLINGTON, VA
Document Type:
Technical Report/Final Report
Publication Date:
2021 Oct 20
Pagination:
11

Security Markings

DOCUMENT & CONTEXTUAL SUMMARY

Distribution Code:
A - Approved For Public Release
Distribution Statement: Public Release.
Copyright: Not Copyrighted

RECORD

Collection: TRECMS
Identifying Numbers
Contract Number(s):
FA9550-20-C-0001
 Subject Terms
Descriptor(s):
neural networks, information systems, artificial intelligence, information science, information processing, air force research laboratories, artificial intelligence software, artificial intelligence computing, deep learning, department of defense, scientific research, topology, algebraic topology, convolutional neural networks, nonlinear dynamics, air force, algorithms
Subject Categories:
Mathematical and Computer Sciences
Creation Date:
2021 Nov 19
Update Date:
2024 Jun 11