Metal-organic frameworks (MOFs) are a class of materials promising for gas adsorption due to their highly tunable nano-porous structures and host-guest interactions. While machine learning (ML) has been leveraged to aid the design or screen of MOFs for different purposes, the needs of big data are not always met, limiting the applicability of ML models trained against small data sets. In this work, we introduce a transfer learning technique to improve the accuracy and applicability of ML models trained with small amount of MOF adsorption data. This technique leverages potentially shareable knowledge from a source task to improve the models on the target tasks. As demonstrations, a deep neural network (DNN) trained on H2 adsorption data with 13,506 MOF structures at 100 bar and 243 K is used as the source task. When transferring knowledge from the source task to H2 adsorption at 100 bar and 130 K (one target task), the predictive accuracy on target task was improved from 0.960 (direct training) to 0.991 (transfer learning). We also tested transfer learning across different gas species (i.e. from H2 to CH4), with predictive accuracy of CH4 adsorption being improved from 0.935 (direct training) to 0.980 (transfer learning). Based on further analysis, transfer learning will always work on the target tasks with low generalizability. However, when transferring the knowledge from the source task to Xe/Kr adsorption, the transfer learning does not improve the predictive accuracy, which is attributed to the lack of common descriptors that is key to the underlying knowledge.