Deep learning is prevalent in many fields and attempts have been made to use it in nonbidirectional mapping problems, such as seismic inversion. These nonbidirectional mapping problems have two special issues, that is, insufficient labels and uncertainty of solution. Therefore, current deep-learning structures are not suitable for handling this kind of problem. A distinctive knowledge-embedded close-looped (KECL) deep-learning framework is developed, tuned to the characteristics of the seismic inverse problem. The KECL deep-learning framework is composed of a reservoir parameter generator (RPG) and a reservoir parameter updater (RPU). The former half-loop is RPG, which takes the seismic data as input to generate the initial reservoir parameters. The latter loop is RPU, which takes the initial parameters as input to output synthetic seismic data. Through the training by well data, the difference between field seismic data and synthetic seismic data modeled by the RPU is used to optimize the RPG and RPU. In this deep-learning framework, knowledge of the Robinson convolutional model is embedded to address the problem of insufficient labels. Furthermore, semisupervised learning is used as prior information to reduce the uncertainty of solution. After the training, with the help of prior geologic information data, the RPU is used to update the initial reservoir parameters generated by RPG for final reservoir parameter inversion. Numerical models and field data are used to test the feasibility of our deep-learning framework. We find that intelligent inversion results using data from one well to train the KECL network are consistent with results using multiple well data. Experiments demonstrate that it is adaptable to situations in which insufficient well data are available and is able to achieve reliable intelligent inversion.