Mobil ve giyilebilir hesaplama için cihaz üzerinde derin öğrenme

Mobile and wearable sensor technologies have gradually extended their usability into a wide range of applications. The amount of collected sensor data can quickly become immense to be processed. The time and resource-consuming computations on such require efficient methods of machine learning and analysis, where deep learning is a promising technique. However, it is challenging to train and run deep learning algorithms on mobile devices due to resource constraints, such as limited battery power, memory, and computation units. In this thesis, we have focused on evaluating the performance of four different deep architectures when optimized with the Tensorflow Lite platform to be deployed on mobile devices in the field of human activity recognition (HAR). We have used two datasets from the literature (WISDM, MobiAct). We have compared the performance of the original models in terms of model accuracy, model size, and resource usages, such as CPU, memory, energy usage, with their optimized versions. As a result of the experiments, we observe that the model sizes and resource consumption were significantly reduced when the models are optimized compared to the original models. Another focus is on personalizing the deep learning model of HAR tasks considering the MobiAct and OpenHAR datasets with a model trained on a larger dataset and with transfer learning that can be fine-tuned on the device with that user if possible. We observed that transfer learning can increase accuracy rate compared to a general model without user-specific data, or a user-specific model trained with small and user-specific data.