BEGIN:VCALENDAR
VERSION:2.0
METHOD:PUBLISH
PRODID:Data::ICal 0.24
BEGIN:VTIMEZONE
TZID:Pacific/Auckland
X-LIC-LOCATION:Pacific/Auckland
BEGIN:DAYLIGHT
DTSTART:19700927T020000
RRULE:FREQ=YEARLY;BYMONTH=9;BYDAY=-1SU
TZNAME:NZDT
TZOFFSETFROM:+1200
TZOFFSETTO:+1300
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:19700405T030000
RRULE:FREQ=YEARLY;BYMONTH=4;BYDAY=1SU
TZNAME:NZST
TZOFFSETFROM:+1300
TZOFFSETTO:+1200
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:ECS Seminars
CONTACT:Shivonne Londt
DESCRIPTION:Manual Convolutional Neural Network (CNN) design for image clas
 sification tasks is costly\, complex\, and time-consuming. Evolutionary Ne
 ural Architecture Search (ENAS) provides an automated alternative in searc
 hing for effective CNNs\, with Genetic Algorithms (GAs) offering flexible 
 search strategies to explore vast\, complex search spaces under well-defin
 ed encodings. This proposal investigates the use of Large Language Models 
 (LLMs) to augment a GA-based ENASframework to improve search performance a
 nd computational efficiency. This research examines five specific areas: t
 he use of LLMs as classifier and repair mechanisms within the GA in a ResN
 et-like search space\, the generalisability of the framework to additional
  CNN architectures\, the introduction of LLM-based dataset-awareness into 
 the search process\, the use of constrained in-algorithm prompt editing vi
 a LLMs to improve search performance\, and the application of agentic augm
 entation to improve speed and overall effectiveness. The expected contribu
 tion is a more adaptive and efficient GA-based evolutionary framework for 
 CNNarchitecture search under practical compute constraints.
DTEND;TZID=Pacific/Auckland:20260421T114500
DTSTAMP:20260420T195412Z
DTSTART;TZID=Pacific/Auckland:20260421T104500
LOCATION:Cotton Club\, Cotton 350
ORGANIZER:Shivonne Londt
SUMMARY:Shivonne Londt - An investigation into whether Large Language Model
 s can improve the performance of evolutionary neural architecture search a
 lgorithms for image classification tasks
UID:seminar_ecs1437_20260417104544
URL:
END:VEVENT
END:VCALENDAR