Vision-based inference systems have recently reached super-human levels thanks to significant progress in deep learning algorithms. However, application of these algorithms on edge devices is challenging due to their limited computational power and limited local information. Alternatively, edge inference systems can utilize the resources available at more capable edge servers to solve the underlying task. Yet, the design of distributed edge inference solutions is challenging, as it requires carefully considering and optimizing multiple factors related to deep learning together with wireless communications.

This thesis studies the design of vision-based edge systems for solving retrieval, classification, and deep neural network parameter delivery tasks under various constraints including the computational, memory, and communication resource limitations. The presented research is based on the recent advances in the field of deep joint source-channel coding (DeepJSCC), which is an alternative to classical, separation-based communication protocols. DeepJSCC simplifies the design of edge systems by introducing an autoencoder neural network, which is trained to map the information source directly to the channel input symbols, and similarly, to map the noisy channel output directly to the reconstructed signal. Such a DeepJSCC autoencoder pair can be further trained with a task-oriented optimization objective, leading to performance gains in the underlying computer vision tasks. For the tasks studied in this thesis, we provide a set of algorithms for achieving improved performance while meeting the communication and computational constraints. Extensive evaluations show that the proposed DeepJSCC approach is an exceptional alternative to the separation-based algorithms, and can play an important role in future generations of intelligent wireless networks.