Machine learning at the mesoscale: a computation-dissipation bottleneck