A modified decentralized model reference adaptive controller is proposed for a continuous time large scale system composed of N interconnected linear subsystems with unknown parameters. In the case where the high frequency gain of the transfer function of each of the isolated subsystems is known, the modified scheme can be used to improve the bad transient behaviour that typically results from poor knowledge of the subsystem parameters. The transient performance improvement, however, has to be traded off against the "size" of the allowable unmodeled interconnections in such a way that as the unmodeled interconnections disappear, the improvement in transient performance can be made arbitrary. Each local adaptive scheme uses "approximate differentiators," and so possible sensor noise amplification is expected to be an additional limitation of the proposed scheme.