Interpretability of artificial neural network models in artificial Intelligence vs. neuroscience

Abstract

Computationally explicit hypotheses of brain function derived from machine learning (ML)-based models have recently revolutionized neuroscience. Despite the unprecedented ability of these artificial neural networks (ANNs) to capture responses in biological neural networks (brains), and our full access to all internal model components (unlike the brain), ANNs are often referred to as black-boxes with limited interpretability. Interpretability, however, is a multi-faceted construct that is used differently across fields. In particular, interpretability, or explainability, efforts in Artificial Intelligence (AI) focus on understanding how different model components contribute to its output (i.e., decision making). In contrast, the neuroscientific interpretability of ANNs requires explicit alignment between model components and neuroscientific constructs (e.g., different brain areas or phenomena, like recurrence or top-down feedback). Given the widespread calls to improve the interpretability of AI systems, we here highlight these different notions of interpretability and argue that the neuroscientific interpretability of ANNs can be pursued in parallel with, but independently from, the ongoing efforts in AI. Certain ML techniques (e.g., deep dream) can be leveraged in both fields, to ask what stimulus optimally activates the specific model features (feature visualization by optimization), or how different features contribute to the model's output (feature attribution). However, without appropriate brain alignment, certain features will remain uninterpretable to neuroscientists.