## Frontmatter | | | | --- | --- | | Authors | [[Umut Güçlü]], [[Marcel van Gerven]] | | Date | 2013/06 | | Source | [[Benelux Conference on Machine Learning]] | | URL | https://benelearn2013.org/pdfs/paper_27.pdf | | Citation | Güçlü, U., & van Gerven, M. (2013). [[Unsupervised learning of features for Bayesian decoding in functional magnetic resonance imaging]]. In _Benelux Conference on Machine Learning_. [[URL](https://benelearn2013.org/pdfs/paper_27.pdf)]. #Conference | ## Abstract Neural decoding is concerned with inferring certain aspects of a stimulus from brain activity. With the recent advent of functional magnetic resonance imaging (fMRI), it has become possible to create a literal picture of a visual stimulus from the human brain. Most conventional decoders are based either on the input space or on a hand-designed feature space. An alternative to hand-designing a feature space is unsupervised feature learning, which has seen much success in computer vision. Here, we present a new decoder, which combines Bayesian inversion of voxel-based encoding models with unsupervised feature learning (independent component analysis). We validated our decoder by reconstructing images of handwritten digits from human brain activity measured using fMRI, with state-of-the-art accuracy. Our results show that the feature space has a substantial effect on the accuracy of the reconstructions, and independent component analysis provides an effective means to learn feature spaces for neural decoding in fMRI. ## PDF ![[Unsupervised learning of features for Bayesian decoding in functional magnetic resonance imaging.pdf]]