Search Papers | Poster Sessions | All Posters
Poster B52 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Algorithms for Neural Networks
John Morrison1,2 (); 1Barnard College, Columbia University, 2Zuckerman Mind Brain Behavior Institute, Columbia University
Algorithms play a central role in cognitive science. They help explain how we perceive, speak, remember, navigate, and decide. But it is unclear what it means say that an artificial or biological neural network "implements" an algorithm. The standard proposal is that a neural network implements an algorithm when it has parts corresponding to the steps of the algorithm. But we haven’t been able to find many such parts, perhaps because neural networks rarely have them. This has led some to deny that neural networks implement algorithms. As an alternative, I propose that a neural network implements an algorithm in virtue of how quickly it learns alternative input-output mappings. This proposal draws on the learning-to-learn literature in psychology and the transfer learning literature in machine learning. I demonstrate that this proposal productively applies to a number of networks and tasks. It is therefore a promising new framework for integrating cognitive science and neuroscience.
Keywords: Neural Networks Algorithms Transfer Learning Mechanistic Interpretability