Multiple-Input Multiple-Output (MIMO) communication using multiple antennas has received significant attention in recent years, both in the academia and industry, as they offer additional spatial dimensions for high-rate and reliable communication, without expending valuable bandwidth. However, exploiting these promised benefits of MIMO systems critically depends on fast and accurate acquisition of Channel State Information (CSI) at the Receiver (CSIR) and the Transmitter (CSIT). In Time Division Duplex (TDD) MIMO systems, where the forward channel and the reverse channel are the same, it is possible to exploit this reciprocity to reduce the overhead involved in acquiring CSI, both in terms of training duration and power. Further, many popular and efficient transmission schemes such as beam forming, spatial multiplexing over dominant channel modes, etc. do not require full CSI at the transmitter. In such cases, it is possible to reduce the Reverse Channel Training (RCT) overhead by only learning the part of the channel that is required for data transmission at the transmitter. In this thesis, we propose and analyze several novel channel-dependent RCT schemes for MIMO systems and analyze their performance in terms of (a) the mean-square error in the channel estimate, (b) lower bounds on the capacity, and (c) the diversity-multiplexing gain tradeoff. We show that the proposed training schemes offer significant performance improvement relative to conventional channel-agnostic RCT schemes. The main take-home messages from this thesis are as follows: • Exploiting CSI while designing the RCT sequence improves the performance. • The training sequence should be designed so as to convey only the part of the CSI required for data transmission by the transmitter. • Power-controlled RCT, when feasible, significantly outperforms fixed power RCT.