One of the central challenges in neuroscience is to understand how the brain turns what we see into patterns of neural activity, and how those patterns can in turn be “read out” to reveal the visual experience. In this thesis, we used a type of artificial intelligence called generative adversarial networks (GANs). These networks are trained to mimic the visual world so convincingly that they can generate entirely new, realistic-looking images from internal codes called latents. Feeding a latent into a GAN produces the image it represents — much like a neural code reflects what a person is seeing at a given moment. By linking brain activity in humans and macaques to these latent codes, we were able to reconstruct pictures of what they were seeing with remarkable accuracy, even though GANs were never trained on brain data. This shows that GANs and the brain rely on surprisingly similar ways of encoding visual information, suggesting that generative principles may be fundamental to how the brain makes sense of the world.
Thirza Dado obtained her bachelor’s degree in Bèta-Gamma (Neurobiology) at the University of Amsterdam, followed by double master’s degrees in Artificial Intelligence and Cognitive Neuroscience (both cum laude) at Radboud University. She is currently a postdoctoral researcher at the Predictive Brain Lab, where she studies belief-based exploration in gamified environments and (un)predictability in the brain.