Cory's Wiki

Introduced in Girshick1), R-CNNs or Regions with Convolutional Neural Networks are an architecture.

The Inception architecture2) combines R-CNNs with strategic use of $1×1$ convolutional layers, and some other techniques (pooling), for a high-performance result.

Is Inception Same as Le et. al.?


How is this the Inception architecture3) different from that of the Le et al architecture4)?

  • both seem to use something like local receptive fields combined with pooling.
  • both address larger-scale hardware and network size tradeoffs
    • Inception “goes deeper”
    • High-level explicitly tries to address large-scale inputs, network size. Inspired directly by cogsci
Sparse Layers