CROSS-LAYER ATTENTION ADAPTATION FOR REAL-TIME NEURAL INFERENCE IN EMBEDDED DEVICES

There is limited processing capabilities, memory, and energy in edge devices that work in real-time setting, which presents a major concern relating to the deployment of deep neural networks on embedded systems. The growing demand of low-latency, energy-efficient inference requires architecture revisions that would not deplete computational resources to maintain accuracy. The proposed paper presents a new framework Cross-Layer Attention Adaptation (CLAA) capable of selectively activating or jumping neural layers in the inference process depending on the complexity of input with the aim to mitigate computation redundancy while sustaining performance. The suggested model utilizes the use of lightweight attention controllers and dynamic gating processes to carry out content-aware layer skipping on a modular network composition of convolutions. It can be scaled on hardware constrained platform like Raspberry Pi 4B and STM32 microcontroller. Some of the primary findings on running benchmark experiments on CIFAR-10 and the Tiny ImageNet would reveal that CLAA minimizes inference latency by 45 percent and energy consumption by more than 36 percent with a similar accuracy as typical full-depth CNNs. The visualisation plot such as pie charts and confusion matrices as well as heat maps confirm the performance and the interpretability of the model. The formulation of the edge AI by the CLAA platform contributes a sustainable and deployable solution to the fledgling landscape of adaptive neural architecture developers as well. It is of great promise in intelligent sensing, mobile vision and timing-constrained automation, especially in applications with tight computation and power requirements.