Toward Approximate Computing for Deep Learning in Embedded Systems: A Systematic Literature Review

The evolution of Deep Learning (DL) algorithms, coupled with the availability of vast amounts of data, has enabled superior accuracy across multiple Artificial Intelligence (AI) tasks. However, such accuracy incurs substantial computational and energy costs, limiting deployment on resource-constrained embedded devices. Approximate Computing has emerged as a promising paradigm to mitigate this challenge by trading off tolerable accuracy loss for significant gains in energy, power, area, and performance. Leveraging the inherent error resilience of DL models, approximation techniques have been extensively explored at multiple layers of the DL computing stack, and several surveys also exist. However, to the best of our knowledge, no prior Systematic Literature Review (SLR) exists that is specifically focused on Approximate Computing for DL in Embedded Systems. This article presents SLR by selecting 51 studies published between January 2019 and May 2025. These studies are organized into four categories corresponding to layers of the DL computing stack, including cross-layer approximations. Furthermore, 12 approximation techniques, 11 benchmark datasets, 26 prominent DL architectures, and three embedded platforms for the energy-efficient DL deployment are summarized, along with the leading tools, frameworks, and optimization/evaluation metrics. Additionally, the academic and industry-wide global adoption has also been highlighted. This SLR reveals that Approximate Multipliers, CIFAR-10 and MNIST datasets, ResNet-50 and LeNet models, and ASIC-based 45 nm technology designs are most frequently targeted, whereas Synopsys DC, PyTorch, and Verilog emerged as the predominant tools and frameworks supporting approximations in DL. Finally, this SLR highlights the current challenges, research gaps, and introduces a conceptual cross-layer approximate DL framework to address them. The findings of this SLR offer a comprehensive reference for the researchers and practitioners to select suitable approximation techniques along with the underlying tools and hardware platforms, aligned with their application requirements.