In local area networks and mobile ad-hoc networks, broadcast-based self-learning is a common mechanism to find packet delivery paths. Self-learning broadcasts the first packet, observes where the returning packet comes from, then creates the corresponding forwarding table entry so that future packets will only need unicast. The main benefits of this mechanism are its simplicity, adaptability, and support of mobility. While the high-level idea of broadcast-based self-learning is straightforward, making the scheme efficient and secure, especially in a data-centric network architecture like Named Data Networking (NDN), requires careful examination. In this paper, we study how broadcast-based self-learning may be applied to NDN networks, point out two major issues: the name-prefix granularity problem and the trust problem, and propose corresponding solutions. We also apply self-learning to switched Ethernet as an example to develop a specific design that can build forwarding tables without any control protocol, recover quickly from link failures, and make use of off-path caches. Simulations are conducted using both real and synthetic traffic to evaluate the performance of the design.