Cyber-Physical Systems (CPS) continues to add more intelligence to various applications, including healthcare. A recent wave of research included the development of pill logging systems that monitor when a pill is extracted. Automatic pill logging systems are an important technology for medication adherence, medication error prevention, and tracing. However, most existing systems do not consider how multi-bottle settings or long-term bottle use can affect their ability to log accurately, forcing frequent re-training/recalibration. In this paper, we present adaptive learning techniques for subject identification across a pair of smart pill bottle systems. We collect inertial signals from subjects taking medication pills and encode the activity signals by transforming them into 2D texture images. Then we use pre-trained Convolutional Neural Network (CNN) models for image-based classification tasks. We evaluate our approach with experiments executed by 10 subjects and achieved improved differentiation capacity over existing models by using deep learning models, modified through domain adaptation and transfer learning.