Ensemble of Deep Architectures. #BigData #DeepLearning #MachineLearning #DataScience
  • Swapout samples from every possible stochastic depth and ResNet architecture, both including and not include dropout!
  • When training the network is constantly mutating as units pick different ways of behaving, but at inference time that network needs to be roughly static so that the same input will always yield the same prediction.
  • Stochastic depth can be thought of as randomly selecting between the outcomes {X, F(X)} for each block, so that every unit in the block returns X or F(X) together.
  • Each box represents a block, and each circle is a unit from the block.
  • A unit that is dropped half the time would (by chance) appear in half randomly generated versions of the network.

Read more


@gp_pulipaka: “Ensemble of Deep Architectures. #BigData #DeepLearning #MachineLearning #DataScience” open tweet »

Show your support

Clapping shows how much you appreciated AI Newsletter’s story.