*Time and date:* 14:00 (CET), Tuesday, 22. 02. 2022.
*Speaker:* Gergő Orbán (Computational Systems Neuroscience Lab, Wigner
Research Centre for Physics)
*Title:* Episodic memory for hypothesis testing
*Zoom link:*
https://ppk-elte-hu.zoom.us/j/95983360984?pwd=OVFvd2lxcUJkZzQ0NG56cmdlYXU5Z…
*Meeting ID:* 959 8336 0984
*Passcode:* 037997
*Abstract:* The ability to evaluate the validity of alternative
hypotheses is one of the cornerstones of learning. This is a major
challenge in artificial learning systems and lack of it has consequences
such as catastrophic interference: the situation where acquired
knowledge is wiped out by new incoming information. Human memory systems
show remarkable resistance to such effects but the computational
principles underlying this ability are not well understood. Using
inspirations from biological memory systems, we propose that a
combination of episodic memory and generative replay can provide support
for close to optimal evaluation of competing hypotheses. We identify two
predictions of this learning framework that concern learning multiple
tasks from observation: specific sensitivity to task structure and
sensitivity to task schedule. We contrast these predictions with data
from human experiments where participants had to learn to perform
multiple tasks with the same set of complex naturalistic stimuli. In
this scenario, our analysis highlights that failure to acquire a new
task coincides with the regime where a normative learning algorithm also
fails to learn.
If you have questions about the event, please contact us via email
(nemecz.zsuzsanna(a)ppk.elte.hu or reka.schvajda(a)ppk.elte.hu).
We look forward to seeing you at the event,
Zsuzsanna Nemecz
Réka Schvajda
organizers
ELTE Department of Cognitive Psychology