NewEvery arXiv paper, its researchers & institutions — mapped.
the archive

#knowledge distillation

6 results
cs.CL2019

Token-Level Ensemble Distillation for Grapheme-to-Phoneme Conversion

Hao Sun, Xu Tan, Jun-Wei Gan +4

The paper introduces a token‑level ensemble distillation method using Transformers to improve grapheme‑to‑phoneme conversion accuracy and reduce model size, leveraging unlabeled da…

#grapheme-to-phoneme conversion#knowledge distillation#token-level distillation#transformer models
cs.CV2019

A Comprehensive Overhaul of Feature Distillation

Byeongho Heo, Jeesoo Kim, Sangdoo Yun +3

The paper proposes a new feature distillation loss that combines a margin ReLU transform, a novel feature position, and a partial L2 distance to improve network compression, achiev…

#feature distillation#network compression#knowledge distillation#image classification
cs.CL2019

Self-Knowledge Distillation in Natural Language Processing

Sangchul Hahn, Heeyoul Choi

The paper introduces a self‑knowledge distillation technique that uses a model’s own soft target probabilities from the embedding space to improve training, and demonstrates its ef…

#knowledge distillation#self-distillation#language modeling#neural machine translation
cs.CV2019

Learning Lightweight Lane Detection CNNs by Self Attention Distillation

Yuenan Hou, Zheng Ma, Chunxiao Liu +1

The paper introduces Self Attention Distillation (SAD), a knowledge‑distillation technique that uses a model’s own attention maps as free supervision to improve lightweight lane‑de…

#lane detection#knowledge distillation#attention mechanisms#lightweight cnn
cs.CV2019

Similarity-Preserving Knowledge Distillation

Frederick Tung, Greg Mori

The paper introduces a knowledge distillation method that trains a student network to preserve the pairwise similarity relationships of activations observed in a teacher network.

#knowledge distillation#representation learning#similarity preservation#model compression
cs.CV2019

Distill-to-Label: Weakly Supervised Instance Labeling Using Knowledge Distillation

Jayaraman J. Thiagarajan, Satyananda Kashyap, Alexandros Karagyris

The paper proposes a weakly supervised method that uses multiple instance learning and knowledge distillation to generate instance-level labels from only image-level annotations, d…

#weak supervision#multiple instance learning#knowledge distillation#histopathology