Core Concepts
CNNは局所性と重み共有のバイアスにより、LCNおよびFCNよりも優れたパフォーマンスを発揮する。
Abstract
Vision tasks are characterized by locality and translation invariance.
CNNs outperform LCNs and FCNs due to the biases of locality and weight sharing.
The DSD classification task models image patches with sparse signal vectors.
CNNs require fewer samples than LCNs and FCNs on the DSD task.
Information theoretic tools are developed for analyzing randomized algorithms.
Stats
CNNは˜O(k + d)サンプルを必要とし、LCNはΩ(kd)サンプルが必要。
LCNは˜O(k(k + d))サンプルを必要とし、FCNはΩ(k2d)サンプルが必要。
Quotes
"Vision tasks are characterized by the properties of locality and translation invariance."
"The superior performance of convolutional neural networks (CNNs) on these tasks is widely attributed to the inductive bias of locality and weight sharing baked into their architecture."