July 21, 2024
Mechanistic Interpretability (MI) promises a path toward fully understanding how neural networks make their predictions. Prior work demonstrates that even when trained to perform simple arithmetic, models can implement a variety of algorithms (sometimes concurrently) depending on initialization and hyperparameters. Does this mean neuron-level interpretability techniques have limited applicability? We argue that high-dimensional neural networks can learn lowdimensional representations of their training data that are useful beyond simply making good predictions. Such representations can be understood through the mechanistic interpretability lens and provide insights that are surprisingly faithful to human-derived domain knowledge. This indicates that such approaches to interpretability can be useful for deriving a new understanding of a problem from models trained to solve it. As a case study, we extract nuclear physics concepts by studying models trained to reproduce nuclear data.
Written by
Ouail Kitouni
Niklas Nolte
Samuel Pérez Díaz
Sokratis Trifinopoulos
Mike Williams
Publisher
icml
Research Topics
Core Machine Learning
August 12, 2024
Arman Zharmagambetov, Yuandong Tian, Aaron Ferber, Bistra Dilkina, Taoan Huang
August 12, 2024
August 09, 2024
Emily Wenger, Eshika Saxena, Mohamed Malhou, Ellie Thieu, Kristin Lauter
August 09, 2024
August 02, 2024
August 02, 2024
July 29, 2024
Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Saketh Rambhatla, Mian Akbar Shah, Xi Yin, Devi Parikh, Ishan Misra
July 29, 2024
Foundational models
Latest news
Foundational models