Підвищення ефективності та точності класифікації одержувачів програми безкоштовного харчування шляхом застосування дистиляції знань в алгоритмі згорткової нейронної мережі

The object of this study is a deep learning-based classification system applied to recipients of the Free Nutritional Food Program, using recipient data as a representation of the eligibility determination process in providing social assistance. The problem to be solved is the high computational complexity of large-capacity convolutional neural networks (CNN) models which, despite their high accuracy, require significant computational resources and are therefore less than optimal for large-scale implementation. To overcome this, this study applies the Knowledge Distillation method, utilizing a large-capacity CNN as the teacher model and a lightweight architecture CNN as the student model through soft label-based knowledge transfer. According to this study, it is shown that the student model generated by distillation is about 90–93% accurate. Also, this figure is very close to that of the teacher model (from 92–95%) and much better than that of a CNN without distillation model (85–88%). This is an improvement since this distillation method can transfer information in the form of richer probabilities than simply hard labels as done do in traditional training. The model proposed in this work has many advantages such as higher accuracy, more compact size, and faster inference times. These features help in making the classification process computationally less intensive. Furthermore, this leads to more efficient memory use and lower energy consumption. These results could be applied in many deep learning classification systems, particularly resource limits devices. Such application is observable in the implementation of Free Nutritious Food Program under real life conditions which requires better accuracy with no loss on efficiency