This thesis develops statistical frameworks that enable direct, cross-individual comparison of spontaneous activity in larval zebrafish. Leveraging whole-brain, single-cell calcium imaging acquired with light-sheet functional microscopy, we use Markov Models and Restricted Boltzmann Machines (RBMs) to build generative, interpretable representations of behavior and neural activity.

First, we characterize spontaneous swimming reorientation with a Hidden Markov Model (HMM) comprising forward, left, and right motor states. The same 3-state architecture also captures the activity of the Anterior Rhombencephalic Turning Region (ARTR), linking neural dynamics to behavior and suggesting that balanced bilateral ARTR activity mediates the forward state.

Next, we introduce two strategies for training RBMs that embed the neural activity of multiple fish into a common latent space. A voxel-based approach trains a single RBM on pooled recordings, while a teacher-student paradigm constrains individual-fish RBMs to shared hidden units. The shared representation allows realistic activity patterns to be transferred between animals, demonstrating that the model captures generic structures conserved across brains.

Finally, we exploit the common latent space to segment spontaneous activity into discrete brain states and quantify their Markovian transition statistics. We find that state-to-state dynamics are remarkably stereotyped across individuals, supporting the hypothesis that spontaneous activity reflects intrinsic priors of neural computation

In conclusion, our results show how probabilistic generative models can bridge individual variability and reveal shared organisational principles of vertebrate brains. The thesis provides a toolbox for cross-subject comparison of spontaneous activity, and underscores the biological interpretability of probabilistic models.