1: \begin{abstract}

2:     How does the human visual cortex encode sensory information? To

3:     address this question, we explore the covariance structure of neural

4:     representations. We perform a cross-decomposition analysis of fMRI

5:     responses to natural images in multiple individuals from the Natural

6:     Scenes Dataset and find that neural representations systematically

7:     exhibit a power-law covariance spectrum over four orders of

8:     magnitude in ranks. This scale-free structure is found in multiple

9:     regions along the visual hierarchy, pointing to the existence of a

10:     generic encoding strategy in visual cortex. We also show that, up to

11:     a rotation, a large ensemble of principal axes of these population

12:     codes are shared across subjects, showing the existence of a

13:     universal high-dimensional representation. This suggests a high

14:     level of convergence in how the human brain learns to represent

15:     natural scenes despite individual differences in neuroanatomy and

16:     experience. We further demonstrate that a spectral approach is

17:     critical for characterizing population codes in their full extent,

18:     and in doing so, we reveal a vast space of uncharted dimensions that

19:     have been out of reach for conventional variance-weighted methods. A

20:     global view of neural representations thus requires embracing their

21:     high-dimensional nature and understanding them statistically rather

22:     than through visual or semantic interpretation of individual

23:     dimensions.

24: \end{abstract}

25: