Reconstructing biological dictionaries: How does neural code translate into behavior?*

The problem of deciphering how low-level patterns (action potentials in the brain, amino acids in a protein, etc.) drive high-level biological features (sensorimotor behavior, enzymatic function) represents the central challenge of modern biology. There are no general methods for reconstructing such dictionaries from small data sets, and for building vocabularies of statistically significant, independent words in them. Here we derive a method for solving this class of problems, which we call the unsupervised Bayesian Ising Approximation (uBIA) and demonstrate its utility in deciphering the motor code in the pre-motor neurons of a songbird. From small data sets, we detect codewords that predict behavior. These words contains arbitrary number of precisely timed spikes, confirming that the motor code in this system is build from precisely timed multi-spike patterns. We show that distinct classes of such words are used preferentially in codes responsible for motor exploration versus exploitation, opening a window on how motor behaviors are controlled by the brain.