Leveraging LSTM Networks for Adaptive Deep Learning in Connected Autonomous Vehicles

In response to the escalating demand for high-performance applications in Connected Autonomous Vehicles (CAVs), there has been an increased effort to accelerate Computer Vision (CV) tasks. Innovations in the field suggest that partitioning Deep Learning (D L) models, which are pivotal for CV tasks, can effectively utilize the computational strengths of both CAVs and proximal edge servers. This is achieved by splitting the DL models into two parts, where execution begins on the CAV, then the intermediate results are offloaded to complete processing on the edge server. Moreover, the precision and response time of these applications vary significantly, necessitating the adoption of early exit strategies. These strategies involve modifying DL models with auxiliary branches that allow for early termination at various stages, enabling the model to exit processing early when application requirements are met. To navigate the trade-offs between accuracy and latency, it is imperative to choose the model's optimal split and exit point that align with the current environmental conditions, such as server data throughput and computational load. To achieve this, we employ a state-of-the-art method that utilizes the Deep Deterministic Policy Gradient (DDPG) algorithm, augmented by Long Short-Term Memory networks. Additionally, to refine the training of the DDPG's actor network, Imitation Learning (IL) is integrated. IL employs a heuristic exhaustive search to retrospectively identify the most effective model partition and exit point.