VALSE2020 重要主题年度进展回顾 (APR)

模型压缩

1. Rethinking Network Pruning
   • Lottery Ticket Hypothesis
   • Rethinking the Value of Network Pruning
     • pruning is an architecture search paradigm
   • Pruning from Scratch (AAAI2020)
2. Data-free Compression
   • Data-free Learning of Student Networks (ICCV2019)
     • GAN 生成训练样本
   • Post trainning 4-bit quentization of concolutional networks for rapid-deployment (NIPS2019)
     • 
   • Towards accurate post-training network quantization via bit-split stitching (ICML2020)
     • 

3. NAS for compact networks

4. NLP compression

1. Hardware-software Co-design
   • 
   • 
   • 

图神经网络

GNN 满足置换不变性，好处是同构的两个节点对应的 topology representation 是一样的。坏处是如果同构的两个节点对应的label不同，这就超出了GNN的表达能力。

Deep GNN 训练存在问题，在图像数据集下有涨点，但是在更general graph setting的数据集下没有performance boost。

场景文字检测与识别

场景文字检测

场景文字识别

端到端场景文字识别

未来展望

元学习

元学习方法论

元学习技术应用

1. few-shot learning
   • 
   • 
   • 
   • 
   • 
2. NAS
3. hyper-parameter learning
   • 
4. continual learning
   • 
5. domain generalization
   • 
6. learning with bias data
   •