机器学习-图像识别-4.pptxVIP

图像识别 - 4计算机视觉问题Sermanet et al, “Integrated Recognition, Localization and Detection using Convolutional Networks”, ICLR 2014Visualizing Deep Learningt-SNE visualization:two images areplaced nearby if theirCNN codes areclose. See more:http://colah.github.io/posts/2015-01-Visualizing-Representations/Practice in CNNDropout vs DropConnectWhat is transfer learning?GPU ComputingGPU - CPU communication is a bottleneck.Implementation details GPUs much faster than CPUs Distributed training is sometimes used Be aware of bottlenecks: CPU / GPU, CPU / disk Low precison makes things faster and still works 3