Work

Exploring the Generalization Capacity of Over-Parameterized Networks

Year: 2023
Type: article
Abstract: Most over-parameterized deep neural networks can generalize to true data. Even over-parameterized networks can learn randomly shuffled bytes in true data, which is far better than just random guessing. In such a scenario where, localized relations are lost to a huge extent, over-parameterized networks show generalization capability. In this paper, multiple models have been trained on various subsets and derivations of CIFAR10 data. Then the models have been trained on true images and true labels, randomly shuffled bytes, and true labels, ordered shuffled bytes and true labels, and randomly generated noise. Shuffling of bytes instead of pixels can distort the image to a great extent, yet over-parameterized models are able to generalize the data. After experimentation, it has been observed that generalization in true pixels is easier in comparison to randomly shuffled bytes data, and generalization in randomly shuffled bytes data is similar to ordered shuffled pixel data. However, memorization is easier and is not affected by any relationships in data.
Language: English
Open Access status: diamond
OpenAlex ID: W7217252990
FWCI: 0.23
Cites: 22
Cited by: 4
Related to: 10