Scene Classification using BNN: Toward a Near Memory Computing design

While Binary Neural Networks (BNNs) offer a promising solution to reduce model complexity for edge-AI embedded inference, they are generally confined to toy or irrelevant benchmarks (e.g., CIFAR-10, ImageNet). Scene classification, which is a better candidate for real-world deployment of BNNs, remains a challenging problem due to high intra-class variability and semantic overlap across categories. It could provide valuable frame understanding for always-on applications performing onsensor inference. BNNs allow for substantial savings in memory and computation, especially if properly mapped onto dedicated hardware such as Near-Memory Computing. Our approach addresses the challenges related to model full binarization by leveraging progressive binarization-aware training and knowledge distillation. We demonstrate that end-to-end BNNs, which are compatible with Near-Memory Computing-based hardware accelerators and dedicated to scene recognition tasks, can effectively capture the spatial and semantic cues required for scene understanding, achieving human-alike accuracy. Our simulations demonstrate that a simple yet efficient model, FBVGG-GRU, and its NMC-compatible adaptation, BSIP-GRU, respectively achieve accuracies of 57.5% and 50.8% on the SUN397 dataset. The model sizes are 5.0 MB and 3.4 MB, respectively, with computational workloads of 67.5 GBOPS and 72.7 GBOPS.