XNL-CNN: An Improved Version Of The NL-CNN Model, For Running With TPU Accelerators And Large Image Datasets

This paper introduces XNL-CNN, an extended version of the previously introduced, compact NL-CNN model. While the NL-CNN was designed with typical Tiny-ML constraints in order to make possible integration of various AI image recognition solutions into low power, low complexity computing platforms, the same stands for its extended version discussed herein. In addition, we consider additional layers to cope with larger image sizes, as found in most actual relevant datasets (e.g. the Cars196 or Flowers104). Moreover, since working with large image sizes requires intensive computational power, TPU accelerators readily available on Kaggle platform were used for training and evaluation of our model. In comparison to other, widely known resources-constrained architectures, such as EfficientNet or MobileNet our model offers the following advantages: i) a larger set of hyper-parameters allowing the user to do better designs (maximal accuracy for minimal resources), adapted for a wide variety of datasets; ii) has a low number of layer primitives, thus making easier their specific design for deployment on various TinyML or EdgeAI platforms, including FPGAS.