BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
X-WR-CALNAME:Yudi Xie Thesis Defense: Neural Network Models of Objects in V
 isual Cognition
X-WR-TIMEZONE:Eastern Time (US & Canada)
BEGIN:VEVENT
DTSTAMP:20260612T141348Z
UID:tag:localist.com\,2008:EventInstance_52566350007876
DTSTART:20260415T180000Z
DTEND:20260415T190000Z
DESCRIPTION:Time: April 15 (Wed)\, 2026\, 2 pm - 3 pm\n\nLocation: Singleto
 n Auditorium (MIT 46-3002)\n\nZoom (if needed): https://mit.zoom.us/j/9937
 5917132\n\nThesis advisor: Prof. James DiCarlo\n\n \n\nAbstract:\n\nObject
 s play an important role in the way we perceive and understand the world t
 hrough vision. My thesis presents three research projects that use task-op
 timized neural network models to gain new insights into how the brain and 
 mind represent\, remember\, and reason about objects from visual inputs. \
 n\n \n\nFirst\, we showed that training convolutional neural networks to e
 stimate object spatial information\, such as object position and pose\, ra
 ther than categories\, can also lead to representations aligned with the p
 rimate ventral visual stream. Furthermore\, spatial tasks and category-tra
 ined networks developed similar representations\, especially in their earl
 y and middle layers. Our results suggest that ventral stream function may 
 be versatile\, and that one should not assume it is optimized for object c
 ategorization alone.\n\n \n\nSecond\, we studied the origin of capacity li
 mitation in working memory by developing image-computable neural network m
 odels. We found that human-like capacity limitations in visual working mem
 ory can be qualitatively explained by models with visual encoders pre-trai
 ned on natural images. In contrast\, models without realistic constraints 
 on sensory encoding did not exhibit human-like memory limitations. Our res
 ults suggest that limitations in visual working memory capacity may be par
 tly due to constraints on realistic sensory encoding.\n\n \n\nThird\, we s
 tudied how humans reason about occluded objects in vision and models that 
 can account for this process. We developed a novel visual reasoning task t
 hat intuitively involves strategies such as imagining compatible shapes an
 d ruling out alternative hypotheses. We found that purely feedforward conv
 olutional neural networks can perform this task well and capture human bia
 ses in some conditions without being explicitly trained to do so. Our find
 ings suggest that purely feedforward models could be powerful\, challengin
 g the view that mechanisms beyond feedforward processing are definitely ne
 eded.\n\n \n\nTaken together\, these projects show various ways in which n
 eural network models can provide new insights into problems in visual cogn
 ition and neuroscience across different levels of analysis.
GEO:42.362302;-71.091766
LOCATION:Building 46\, 3002
SUMMARY:Yudi Xie Thesis Defense: Neural Network Models of Objects in Visual
  Cognition
URL;VALUE=URI:https://calendar.mit.edu/event/yudi-xie-thesis-defense-neural
 -network-models-of-objects-in-visual-cognition
CATEGORIES:Thesis defense
END:VEVENT
END:VCALENDAR