FRINK Linked Data Fragments server

Matches in SemOpenAlex for { <https://semopenalex.org/work/W4380479929> ?p ?o ?g. }

• W4380479929 endingPage "63" @default.
• W4380479929 startingPage "33" @default.
• W4380479929 abstract "Feature-based Knowledge Distillation (FKD) is a method for guiding the activations at the intermediate layers of a Convolutional Neural NetworkConvolutional Neural Network (CNN) (CNN) during training. It has recently gained significant popularity as either a standalone or complementary method of Knowledge Distillation (KD)Knowledge distillation. Most techniques however, handle the teacher-to-student knowledge transfer in a statistical or probabilistic manner. In this chapter, we propose a family of efficient loss functions that aim to transfer the geometry of activations from intermediate layers of a teacher CNNConvolutional Neural Network (CNN) to the activations of a student CNN model at matching spatial resolutions through a novel Feature-based Knowledge Distillation (FKD) method. We discuss the challenges of geometric methods, provide connections with manifold-to-manifold comparison and regularization techniques, and draw some interesting connections to the field of random graphs. Experiments on benchmark tasks show evidence that by focusing on replicating the feature relationships only across local neighborhoods, results in better performance. Furthermore, the definition of neighborhoods important for sufficient performance, with neighborhoods defined over parsimonious graphs such as the Minimal Spanning Tree achieving better results that standard kNN rule. Finally, we present a case study on data-free Knowledge Distillation in the field of offline handwritten signature verification. The case study demonstrates a way to harness knowledge from an expert CNNConvolutional Neural Network (CNN) model to enhance the training of a new model with different architecture, using only external (task-irrelevant) data. Results indicate that both geometric FKD and its combination with standard KDKnowledge distillation techniques can effectively create new models with better performance than the expert teacher model." @default.
• W4380479929 created "2023-06-14" @default.
• W4380479929 creator A5008187790 @default.
• W4380479929 creator A5056682392 @default.
• W4380479929 date "2023-01-01" @default.
• W4380479929 modified "2023-10-16" @default.
• W4380479929 title "A Geometric Perspective on Feature-Based Distillation" @default.
• W4380479929 cites W1967277558 @default.
• W4380479929 cites W1989368986 @default.
• W4380479929 cites W2013188364 @default.
• W4380479929 cites W2048916794 @default.
• W4380479929 cites W2050995497 @default.
• W4380479929 cites W2052330882 @default.
• W4380479929 cites W2052468365 @default.
• W4380479929 cites W2068396114 @default.
• W4380479929 cites W2077887717 @default.
• W4380479929 cites W2087399108 @default.
• W4380479929 cites W2090504921 @default.
• W4380479929 cites W2104339857 @default.
• W4380479929 cites W2110599581 @default.
• W4380479929 cites W2126017757 @default.
• W4380479929 cites W2194775991 @default.
• W4380479929 cites W2213490786 @default.
• W4380479929 cites W2344608894 @default.
• W4380479929 cites W2614376571 @default.
• W4380479929 cites W2731516819 @default.
• W4380479929 cites W2739879705 @default.
• W4380479929 cites W2884036341 @default.
• W4380479929 cites W2887783173 @default.
• W4380479929 cites W2898714651 @default.
• W4380479929 cites W2904826325 @default.
• W4380479929 cites W2949658730 @default.
• W4380479929 cites W2963132249 @default.
• W4380479929 cites W2963140444 @default.
• W4380479929 cites W2965031610 @default.
• W4380479929 cites W3010510250 @default.
• W4380479929 cites W3015858876 @default.
• W4380479929 cites W3034368386 @default.
• W4380479929 cites W3035476138 @default.
• W4380479929 cites W3088421935 @default.
• W4380479929 cites W3093859587 @default.
• W4380479929 cites W3094001968 @default.
• W4380479929 cites W3101813758 @default.
• W4380479929 cites W3102713119 @default.
• W4380479929 cites W3127300035 @default.
• W4380479929 cites W3158763993 @default.
• W4380479929 cites W3174102142 @default.
• W4380479929 cites W3210157601 @default.
• W4380479929 cites W4200444732 @default.
• W4380479929 cites W4220784493 @default.
• W4380479929 cites W4289600407 @default.
• W4380479929 doi "https://doi.org/10.1007/978-3-031-32095-8_2" @default.
• W4380479929 hasPublicationYear "2023" @default.
• W4380479929 type Work @default.
• W4380479929 citedByCount "0" @default.
• W4380479929 crossrefType "book-chapter" @default.
• W4380479929 hasAuthorship W4380479929A5008187790 @default.
• W4380479929 hasAuthorship W4380479929A5056682392 @default.
• W4380479929 hasConcept C119857082 @default.
• W4380479929 hasConcept C138885662 @default.
• W4380479929 hasConcept C153180895 @default.
• W4380479929 hasConcept C154945302 @default.
• W4380479929 hasConcept C178790620 @default.
• W4380479929 hasConcept C185592680 @default.
• W4380479929 hasConcept C202444582 @default.
• W4380479929 hasConcept C204030448 @default.
• W4380479929 hasConcept C2776401178 @default.
• W4380479929 hasConcept C33923547 @default.
• W4380479929 hasConcept C41008148 @default.
• W4380479929 hasConcept C41895202 @default.
• W4380479929 hasConcept C49937458 @default.
• W4380479929 hasConcept C50644808 @default.
• W4380479929 hasConcept C81363708 @default.
• W4380479929 hasConcept C9652623 @default.
• W4380479929 hasConceptScore W4380479929C119857082 @default.
• W4380479929 hasConceptScore W4380479929C138885662 @default.
• W4380479929 hasConceptScore W4380479929C153180895 @default.
• W4380479929 hasConceptScore W4380479929C154945302 @default.
• W4380479929 hasConceptScore W4380479929C178790620 @default.
• W4380479929 hasConceptScore W4380479929C185592680 @default.
• W4380479929 hasConceptScore W4380479929C202444582 @default.
• W4380479929 hasConceptScore W4380479929C204030448 @default.
• W4380479929 hasConceptScore W4380479929C2776401178 @default.
• W4380479929 hasConceptScore W4380479929C33923547 @default.
• W4380479929 hasConceptScore W4380479929C41008148 @default.
• W4380479929 hasConceptScore W4380479929C41895202 @default.
• W4380479929 hasConceptScore W4380479929C49937458 @default.
• W4380479929 hasConceptScore W4380479929C50644808 @default.
• W4380479929 hasConceptScore W4380479929C81363708 @default.
• W4380479929 hasConceptScore W4380479929C9652623 @default.
• W4380479929 hasLocation W43804799291 @default.
• W4380479929 hasOpenAccess W4380479929 @default.
• W4380479929 hasPrimaryLocation W43804799291 @default.
• W4380479929 hasRelatedWork W2175746458 @default.
• W4380479929 hasRelatedWork W2732542196 @default.
• W4380479929 hasRelatedWork W2738221750 @default.
• W4380479929 hasRelatedWork W2760085659 @default.
• W4380479929 hasRelatedWork W2767651786 @default.

• next