Token Labeling: Training a 85.4% Top-1 Accuracy Vision Transformer with 56M Parameters on ImageNet.
Zihang JiangQibin HouLi YuanDaquan ZhouXiaojie JinAnran WangJiashi FengPublished in: CoRR (2021)
Keyphrases
- computational cost
- prediction accuracy
- training set
- real time
- window size
- vision system
- high accuracy
- computer vision
- labeling effort
- test set
- active learning
- image processing
- maximum likelihood
- expectation maximization
- parameter estimation
- error rate
- training process
- training and testing data
- training speed
- labeled data for training
- image collections
- visual features
- object detection
- image segmentation
- learning algorithm
- machine learning