Karen Simonyan

Cofounder, Inflection AI

Search every verified Karen Simonyan interview, podcast appearance, and on-the-record quote — each transcript cross-checked by AI and human review to confirm speaker identity. At the 2015 International Conference on Learning Representations (ICLR), Karen Simonyan presented joint work with Andrew Zisserman on very deep convolutional networks for large-scale image recognition. In the talk, he described the main objective of the research as determining how far performance could be improved by increasing the depth of image classification networks. He noted that the networks evaluated shared the same architecture design, differing only in depth, and that the models were made publicly available for research use. He also discussed findings that using multiple scales at training and test time was beneficial, and that dense evaluation and multiple crops yielded compatible results that were complementary when combined. Simonyan also addressed the relationship between network depth and parameter structure, stating that the number of parameters in their architecture was roughly the same as in AlexNet, but that the parameters were structured to create a very deep network with fewer parameters per layer and more layers overall. He attributed the team's second-place finish in the 2014 ImageNet classification challenge to these design choices, while noting that further improvements in depth were limited by training initialization methods at the time, which he said were later found to be unnecessary with appropriately scaled random initialization.