Motor behavior appears to be disorganized during early stages of development and is thought to be initially purely explorative. According to current theories, such exploration is essential for learning sensorimotor transformations. We combine data analysis and modeling to study the central neuronal mechanism that underlies explorative behaviors, focusing on babbling-like behaviors in human and non-human vocal learners. We quantify the temporal structure of babbling in four vocal learners with different levels of complexity in their adult vocalizations: zebra finches, swamp sparrows, canaries and human babies. We show that: 1) In all these learners, gesture duration distributions during early babbling are well fitted by a decaying exponential (in line with previous reports in zebra finches; ) ; 2) interspecies differences in the temporal features of early babbling are to a large extent accounted for by time rescaling. These findings point to the existence of a central mechanism underlying explorative behavior which is common to all these learners and therefore robust with respect to anatomical and physiological differences between individuals and species. We investigate possible mechanisms for motor exploration in a theoretical model of the neuronal circuit that activates the effectors producing babbling in songbirds. It comprises a premotor and a motor network representing the avian cortical-like areas LMAN and RA, each consisting of a large number of excitatory and inhibitory spiking neurons. The premotor network projects to the motor network which in turn activates a small number of effectors as is the case in songbird anatomy. We argue that requiring the circuit to autonomously and robustly generate random activation of the effectors constrains its architecture and dynamics strongly. We show that this implies that: 1) the premotor as well as the motor area are recurrent networks operating in a regime where excitation and inhibition are balanced [2,3,4] and 2) the feedforward projections from the former to the latter and then to the effectors are topographic. Under these conditions, the motor network exhibits temporally irregular firing with substantial correlations between neurons that activate the same effector. Importantly, correlations emerge from the recurrent dynamics of the circuit without any fine- tuning of parameters. When connected to a non linear model of a syrinx , the circuit generates explorative behavior with statistics similar to those exhibited in the data. Finally, we validate our theory by testing its predictions regarding the spatiotemporal patterns of activity in LMAN and RA in neuronal recordings in singing finches.
Work conducted in the framework of the France Israel Laboratory of Neuroscience.
 Aronov, D., Andalman, A. S., & Fee, M. S. Science 320(5876):630-634 (2008).
van Vreeswijk, C., & Sompolinsky, H. Science 274(5293):1724-1726 (1996).
van Vreeswijk, C., & Sompolinsky, H. Neural computation 10(6):1321-1371 (1998).
 Renart, A., de la Rocha, J., Bartho, P., Hollender, L., Parga, N., Reyes, A., & Harris, K. D. J. Science. 327(5965):587-590. (2010).
 Amador, A., Perl, Y. S., Mindlin, G. B., & Margoliash, D. Nature. 495(7439):59-64. (2013).