Multi scale dense networks

MULTI-SCALE DENSE NETWORKS
FOR RESOURCE EFFICIENT IMAGE CLASSIFICATION
Gao Huang et al.,ICLR 2018(Oral)
Motivation
• Two Settings:
• Anytime classification
• Budgeted batch classification
• A test image
• “easy” vs. “hard”
• shallow model vs. deep model
• Training multiple classifiers with varying
resource demands, which we adaptively
apply during test time.
Motivation
• Develop CNNs
• “slice” the computation and process these slices one-by-one, stopping the
evaluation once the CPU time is depleted or the classification sufficiently
certain (through “early exits”)
classifier
…feature
Motivation
• Problems
• The lack of coarse-level features of early-exit classifiers
• Early classifiers interfere with later classifier
Multi-Scale DenseNet (MSDNet)
• Solutions
• Multi-scale feature maps
• Dense connectivity
Multi-Scale DenseNet (MSDNet)
• Architecture
• First Layer
• Subsequent layers
• Classifiers : conv + conv + average pooling + linear layer
• Loss Functions : cross entropy loss
Multi-Scale DenseNet (MSDNet)
• Network reduction
S – i + 1
Multi-Scale DenseNet (MSDNet)
• Two Settings:
• Anytime classification – output the most recent prediction until the budget is
exhausted
• Budgeted batch classification – exits after classifier fk if its prediction
confidence exceeds a pre-determined threshold
Multi-Scale DenseNet (MSDNet)
• Experiments
• Anytime Prediction
Multi-Scale DenseNet (MSDNet)
• Experiments
• Budged batch classification
Multi-Scale DenseNet (MSDNet)
• Visualization
1 of 11

More Related Content

Multi scale dense networks

Editor's Notes

  1. Setting 1: 测试样本任意时间内输出预测结果; Setting 2:一个较大样本集共享固定的计算资源; 测试样本易于识别或计算资源有限, 使用小模型; 难以识别或计算资源充裕, 使用大模型;
  2. Problem 1: 提取最后一层特征直接用于分类器学习, 而浅层特征并没有直接用于分类器学习; Problem 2: 网络不同层的特征尺度不同。 通常情况下, 深层网络的浅层用于提取fine scale low-level特征,而后边的层用于提取coarse scale特征;