WMREPS Hidden brain states underlying efficient representations in working memory
Zusammenfassung der Projektergebnisse
Working memory uncouples the employment of information for ongoing task demands from the availability of external stimuli and is thus critical for adaptive behavior. Its flexibility is further enhanced by prioritization processes where executive control regions selectively highlight task-relevant stored contents in the service of behavioral demands. How this executive control takes place, and in which representational format the prioritized representation is activated in the brain, is currently a matter of intense research. VWM prioritization has been proposed to require distinct coding schemes and representational transformations in the PFC. The main objective of the WMREPS project was to delve into the specifics of these transformations and investigate which aspects of VWM representations are highlighted during VWM prioritization. From a theoretical perspective, the project aimed to uncover the computational architectures that support the representational transformations underlying VWM prioritization, and the correspondent neurophysiological mechanisms in the human brain. An innovative multi-item visual working memory task was developed and conducted with different groups of subjects, including healthy human participants and patients implanted with iEEG electrodes. In the task, three items were presented in a sequence and a retro-cue prompted subjects to maintain only one of them or all of them. Advanced methodological approaches, including RSA and DNN models, were applied to the iEEG data to investigate the formation of category-specific representations and track their transformation during VWM maintenance and prioritization. The project yielded several important results. A robust presence of category-specific information was observed during encoding in the oscillatory time-frequency patterns of activity recorded in the VVS and the PFC. During maintenance, this information re-occurred in VVS but not in PFC – suggesting a transformation of PFC representations from encoding to maintenance. Findings revealed that representations during the maintenance period matched representations in deep layers of recurrent but not feedforward DNNs, in both VVS and PFC, emphasizing the critical role of recurrent computations for VWM. While recurrent DNN representations matched PFC representations in the beta band following the retro-cue, they corresponded to VVS representations in a lower theta-alpha frequency range (3-14Hz) towards the end of the maintenance period. These results could be replicated in recurrent DNNs with two different architectures and using two different training sets. Critically, the strongest matching was observed specifically in the deep layers of the recurrent models, highlighting the abstract representational format of prioritized content in the PFC. In summary, the WMREPS project delved into the neural mechanisms of VWM maintenance and prioritization, demonstrating that VWM prioritization relies on distinct representational transformations and formats in the VVS and the PFC. While some of the originally planned analysis and approaches had to be adopted due to the COVID pandemic, the project has produced important knowledge regarding the nature of VWM representations, their computational principles and neurophysiological substrate. The outcomes of the project have been published and disseminated through various presentations at international conferences and workshops.
Projektbezogene Publikationen (Auswahl)
-
Concurrent maintenance of both veridical and transformed working memory representations within unique coding schemes.
Kandemir, Güven; Wolff, Michael J.; Karabay, Aytaç; Stokes, Mark G.; Axmacher, Nikolai & Akyürek, Elkan G.
-
How can we track cognitive representations with Deep Neural Networks and intracranial EEG? In Intracranial EEG. A Guide for Cognitive Neuroscientists (Springer). ISBN 978-3-031-20909-3.
Pacheco Estefan, D.
-
Top-down control of multi-item working memory representations. Conference presentation at the Cognitive Neuroscience conference (CNS), April 23-26, 2022, San Francisco, California, USA.
Pacheco Estefan, D., Fellner, M.C., Zhang, H., Kunz, L., Reinacher, P., Roy, C., Schulze-Bonhage, A., Yang, L., Wang, S., Liu, J., Xue, G. & Axmacher, N.
-
Top-down control of multi-item working memory representations. Poster presentation at the Working memory engram meeting, May 30-31, 2022, Groningen, the Netherlands.
Pacheco Estefan, D., Fellner, M.C., Zhang, H., Kunz, L., Reinacher, P., Roy, C., Schulze-Bonhage, A., Yang, L., Wang, S., Liu, J., Xue, G. & Axmacher, N.
-
Maintenance of colour memoranda in activity-quiescent working memory states: Evidence from impulse perturbation.
Kandemir, Güven; Wilhelm, Sophia A.; Axmacher, Nikolai & Akyürek, Elkan G.
