Soft strain sensors are becoming increasingly popular for obtaining tactile information in soft robotic applications. Diverse technological solutions are being investigated to design these sensors. Simultaneously, new methods for modeling these sensor are being proposed due to their highly nonlinear, time varying properties. Among them, machine learning based approaches, particularly using dynamic recurrent neural networks look the most promising. However, these complex networks have large number of free parameters to be tuned, making it difficult to apply them for real-world applications. This paper introduces the concept of transfer learning for modelling soft strain sensors, which allows us to utilize information learned in one task to be applied to another task. We demonstrate this technique on a passive anthropomorphic finger with embedded strain sensors used for two regression tasks. We show how the transfer learning approach can drastically reduce the number of free parameters to be tuned for learning new skills. This work is an important step towards scaling of sensor networks (algorithm-wise) and for using soft sensor data for high-level control tasks.