Ultra-reliable and low-latency communication (URLLC) is a new service category to accommodate emerging services requiring low end-to-end latency. In the 4G LTE/5G NR, it is very difficult to satisfy the URLLC requirements since multiple OFDM symbols are processed as a bundle. In this paper, we propose a novel low-latency packet transmission scheme, referred to as ultra-mini slot transmission (UMST), suitable for the short packet transmission in URLLC scenario. Key idea of the proposed scheme is to transform the transmit information into subcarrier positions and then decode it using a small amount of time-domain received samples. In particular, in the UMST decoding, we put forth an entirely different approach based on a deep neural network (DNN). From the numerical evaluations on realistic channel models, we demonstrate that the UMST scheme outperforms the conventional transmission schemes in terms of the block error rate (BLER) and transmission latency.