神经网络其实并不需要那么深!普林斯顿大学&英特尔提出ParNet,12层的网络就能达到80%以上的准确率!
论文链接:https://arxiv.org/abs/2110.07641
项目链接:https://github.com/imankgoyal/NonDeepNetworks

01

02

2.1. ParNet Block

2.2. 下采样和融合块

2.3. 网络架构

2.4. Scaling ParNet

03
3.1. Experiments on ImageNet



3.2. Boosting Performance

3.3. Experiments on MS-COCO

3.4. Experiments on CIFAR

3.5. Ablation Studies


3.6. ParNet vs. Ensembles

3.7. Scaling ParNet

04

END
赞 (0)