| Title | Analysis of loss function for binary classification problem in deep learning / |
|---|---|
| Translation of Title | Tikslo funkcijos analizė binarinio klasifikavimo uždaviniui spręsti naudojant gilųjį mokymąsi. |
| Authors | Kondrat, Nikolaj |
| Full Text |
|
| Pages | 87 |
| Keywords [eng] | Gilusis mokymasis, tikslo funkcija, binarinis klasifikavimas, veidų verifikacija, ROC kreivė, DET kreivė Deep learning, loss function, binary classification, face verification, receiver operator characteristic curve, ROC, detection error tradeoff curve, DET |
| Abstract [eng] | The thesis analyzes the applicability of different deep learning loss functions for binary classification problems. Misclassification error-rate based generic loss functions are not directly adapted for optimization of binary classification specific metrics. Consequently, different binary classification problems in various scenarios have custom evaluation requirements that are not being addressed. To bridge this gap, an investigation of dependencies between loss function form and its performance under different evaluation constraints is performed. Biometric face verification is chosen as a problem domain. Three commonly used loss functions are analyzed using three tasks from this domain. Based on investigation results, practical advice pack how to choose appropriate deep learning loss function for differently constrained binary classification problems from biometric face verification domain is presented. |
| Dissertation Institution | Vilniaus universitetas. |
| Type | Master thesis |
| Language | English |
| Publication date | 2019 |