BEGIN:VCALENDAR
VERSION:2.0
PRODID:IEEE vTools.Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:America/Chicago
BEGIN:DAYLIGHT
DTSTART:20270314T030000
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3
TZNAME:CDT
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20261101T010000
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11
TZNAME:CST
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20261112T195013Z
UID:A7006721-47EF-451C-87C2-808665A54D84
DTSTART;TZID=America/Chicago:20261120T173000
DTEND;TZID=America/Chicago:20261120T183000
DESCRIPTION:The inability to establish appropriate assurance for increasing
 ly autonomous systems presents a critical barrier to their safe deployment
  in aviation and other safety-critical domains. This challenge drives deve
 lopment costs uneconomically high\, delays adoption\, and creates unknown 
 and unmanageable risks. This talk explores the fundamental question: What 
 constitutes sufficient evidence to substantiate safety claims\nfor increas
 ingly autonomous systems performing safety-critical functions? In this tal
 k we establish foundational concepts of safety assurance\, distinguishing 
 between design safety and operational safety\, and examine why assurance r
 emains fundamentally difficult. The core of the talk explores two compleme
 ntary research thrusts. First\, we investigate assurance of machine learni
 ng systems through a case study of AssistNet\, a deep learning model for r
 unway and taxiway detection. Using formal methods tools\, we demonstrate h
 ow formal bound propagation and branch-and-bound algorithms can provide ma
 thematical guarantees of neural network robustness against defined perturb
 ations within an operational design domain. This work reveals both the pro
 mise and current limitations of formal verification for safety-critical co
 mputer vision applications. Second\, we examine how machine learning can a
 ccelerate assurance processes themselves. We present machine-learning enha
 ncements that can democratize formal theorem proving by translating natura
 l language specifications into formal proofs\, auto-suggesting proof tacti
 cs\, and performing autonomous verification attempts. The presentation con
 cludes by postulating paths forward to develop standardized\, scalable cer
 tification pathways. While no single technique suffices\, a combination of
  means\, methods\, tools and techniques can help in the assurance of incre
 asingly autonomous aviation systems\, thereby maintaining the rigorous saf
 ety standards the public demands while enabling innovation.\n\nSpeaker(s):
  Dr. Natasha Neogi\, \n\nVirtual: https://events.vtools.ieee.org/m/591160
LOCATION:Virtual: https://events.vtools.ieee.org/m/591160
ORGANIZER:mvvrmkr@gmail.com
SEQUENCE:33
SUMMARY:Assurance Methods for Increasingly Autonomous Systems
URL;VALUE=URI:https://events.vtools.ieee.org/m/591160
X-ALT-DESC:Description: &lt;br /&gt;&lt;p&gt;The inability to establish appropriate ass
 urance for increasingly autonomous systems&amp;nbsp\;presents a critical barri
 er to their safe deployment in aviation and other safety-critical&amp;nbsp\;do
 mains. This challenge drives development costs uneconomically high\, delay
 s&amp;nbsp\;adoption\, and creates unknown and unmanageable risks. This talk e
 xplores the&amp;nbsp\;fundamental question: What constitutes sufficient eviden
 ce to substantiate safety claims&lt;br&gt;for increasingly autonomous systems pe
 rforming safety-critical functions? In this talk we establish foundational
  concepts of safety assurance\, distinguishing between design safety and o
 perational safety\, and examine why assurance remains fundamentally diffic
 ult. The core of the talk explores two complementary research thrusts. Fir
 st\, we investigate assurance of machine learning systems through a case s
 tudy of AssistNet\, a deep learning model for runway and taxiway detection
 . Using formal methods tools\, we demonstrate how formal bound propagation
  and branch-and-bound algorithms can provide mathematical guarantees of ne
 ural network robustness against defined perturbations within an operationa
 l design domain. This work reveals both the promise and current limitation
 s of formal verification for safety-critical computer vision applications.
  Second\, we examine how machine learning can accelerate assurance process
 es themselves. We present machine-learning enhancements that&amp;nbsp\;can dem
 ocratize formal theorem proving by translating natural language specificat
 ions&amp;nbsp\;into formal proofs\, auto-suggesting proof tactics\, and perfor
 ming autonomous&amp;nbsp\;verification attempts. The presentation concludes by
  postulating paths forward to&amp;nbsp\;develop standardized\, scalable certif
 ication pathways.&amp;nbsp\;While no single technique suffices\, a combination
  of means\, methods\, tools and&amp;nbsp\;techniques can help in the assurance
  of increasingly autonomous aviation systems\,&amp;nbsp\;thereby maintaining t
 he rigorous safety standards the public demands while enabling&amp;nbsp\;innov
 ation.&lt;/p&gt;
END:VEVENT
END:VCALENDAR

