[home] [Personal Program] [Help]

---

Session: Poster Session 1 (Even numbers)
Session starts: Thursday 26 January, 16:00
Presentation starts: 16:00

---

Thirza Dado (Radboud University, Donders Institute for Brain, Cognition and Behaviour)
Yağmur Güçlütürk (Radboud University, Donders Institute for Brain, Cognition and Behaviour)
Marcel van Gerven (Radboud University, Donders Institute for Brain, Cognition and Behaviour)
Umut Güçlü (Radboud University, Donders Institute for Brain, Cognition and Behaviour)

Abstract:
The field of neural decoding seeks to find which information about a perceived sensory stimulus is present in and can be retrieved from recorded brain activity. The adoption of generative adversarial networks (GANs) in neural decoding has been fruitful in that they effectively model the “synthesis” operation from feature space to real-world data samples. As such, it is possible to reconstruct GAN-synthesized face stimuli that were presented to participants based on their recorded neural activations. The close resemblance between the stimuli and their reconstructions from brain data indicates that GAN latent- and neural representations represent the images similarly. Here, we present how to more closely approximate the presented stimuli as perceived by two participants in the MRI scanner by hyper-aligning and -decoding their fMRI recordings. This approach enabled the training of a single general decoder model that captures the shared neural information of the two participants to more accurately predict the input stimulus.