Task-guided cross-subject latent alignment: a multi-encoder-decoder VAE
Researchers have developed a novel Multi-Encoder-Decoder Variational Autoencoder (MED-VAE) that can align neural activity across different subjects without requiring shared stimuli. This method anchors representations to a pretrained artificial neural network, enabling the discovery of shared computational principles and generalizable decoders. Experiments using the Natural Scenes Dataset demonstrated that MED-VAE creates semantically organized common latent spaces, outperforming traditional methods in cross-subject alignment and robust generalization to novel stimuli. AI