视觉更好地分析深度卷积神经网络.pdf

Towards Bet nalysis of Deep Convolutional Neura works Mengchen Liu, Jiaxin Shi, Zhen Li, Chongxuan Li, Jun Zhu, Shixia Liu Tsinghua University Background • Deep convolutional neura works (CNNs) have demonstrated significant improvements on many tasks, such as image classification and the game of Go DeepMind challenge match One key factor: two deep CNNs 4-1 Mar 2016 Why do the neural networks work? L Sedol (9p) Beats Beats Convolution Nature match 4-1 Oct 2015 Fan Hui (2p) work Valu work A Visual ytics System CNNVis Challenge • The size of a CNN is typically large – Tens or hundreds of layers (depth) – Thousands of neurons in each layer (width) – Millions of connections between neurons • Many functional components – Their values and roles are not well understood Our Contribution • A visual ytics system – Understanding – Diagnosis – Refinement • A hybrid visualization – Rectangle packing – Matrix visualization – Biclustering-based edge bundling CNNVis Overview Visual Encoding A neuron cluster A neuron Visual Encoding Learned features Activations Visual Encoding Connections Edge b