We study the impact of network architecture on the efficiency of information transmission and dynamics of learning in medium-sized networks using laboratory experiments. Subjects are incentivized to guess the correct state of nature based on (i) a private noisy signal they get before the game begins and (ii) the past guesses of their immediate neighbors. We show that networks in which a single individual observes everyone are performing surprisingly badly when the distribution of private signals unambiguously favors one action over the other.  This observation is at odds with Bayesian theory according to which information aggregation on such networks should be simple and quick. We trace this failure to two behavioral frictions - (i) excess reliance on private signals and (ii) under imitation of immediate neighbors who have better information. We also show that these two biases are interdependent. We explore other network structures that do not have a global aggregator and show that these behavioral frictions are present in all networks and all positions, but their effect on overall aggregation is different due to architectural differences. Finally, we find evidence for non-monotonicity in the quality of aggregation with respect to the quantity of information.