This event will be in-person ONLY in Wu and Chen Auditorium.

ABSTRACT

Black box methods such as the Turing Test are no longer adequate for evaluating AI models. Quantifying the structure and similarity of high-dimensional neural representations are essential for better understanding and training of large neural network models. Statistical insights can be gained by analyzing the geometrical structure of these representations as they are reformatted by neural network hierarchies. Unfortunately, conventional analysis based upon covariance matrices yield biased estimators under realistic constraints of finite data. I will present some recent advances in developing efficient estimators that address these limitations. Our results highlight the importance of developing principled statistical tools to analyze neural representations.