Auto Curation on FaceNet Embeddings with Gamma and Gaussian Distribution to Predict Model Performance in Actual Industrial Deployment

Abstract

Many AI applications, such as face recognition [1] and NLP, rely heavily on data embedding as an intermediate representation on which further processing is made. However few of these applications gain insights to such intermediate representation, and thus have difficulties in data analytic or designing efficient models, or both. The resulting models accordingly designed are thus hard to analyze for performance tuning and optimization. We deeply dived into the embedding of FaceNet in an actual industrial deployed site, and propose a closed-loop solution with data representation, data curation, data modeling on these intermediate data, as to do performance prediction for 1:1 and 1:N scenarios [2]. Our results shows our prediction of the model, in the range of interest of application, achieved 0.4% error in predicting True Positive Rates, and 2.8% error in predicting False Positive Rates.