Коротаев Кирилл Алексеевич
Using TMS to Identify Links Between Deep Neural Networks and Visual Cortex
Когнитивные науки и технологии: от нейрона к познанию
CNN revolutionized machine learning by becoming state of art in object recognition. Aside from being very accurate, these models have claimed to share similarities with human visual cortex. But the question of whether the same computational mechanisms are used in visual cortex remains unclear. Here, instead of simple accuracy comparison between human and model, we investigate if CNN commit same errors as our participants during identification task. We trained a CNN to discriminate between same and different pairs of faces and houses and compared performance to experimental data. We established that CNN could reach the same accuracy as our participants, and that it was capable of reproducing FIE. Next, we applied TMS to our subjects during the same task over OFA and OPA. Finally, we attempt to mimic TMS effects onto CNN by knocking out weights in the medium convlayer. We have proven that it is possible to knockout weights that are exclusively sensitive to either faces or scenes. Our results show that it is possible for CNN to reproduce human errors in an identification task and that knocking out weights in specific CNN layers can reproduce artificial lesions from TMS.