For deep learning detection networks in the multiple-input-multiple-output (MIMO) channel, deepening the network does not significantly improve performance beyond a certain number of layers. In this letter, we propose a parallel detection network (PDN) that consists of several deep learning detection networks in parallel without connection. By designing a specific loss function and reducing similarity between detection networks, the PDN obtains a considerable diversity effect. The performance of the PDN improves significantly as the number of parallel detection networks increases in time-varying MIMO channels. This is superior to the existing deep learning detection networks, in both performance and complexity.