Systematic Perturbation of an Artificial Neural Network: A Step Towards Quantifying Causal Contributions in The Brain

Abstract

Lesion inference analysis is a fundamental approach for characterizing the causal contributions of neural elements to brain function. Historically, it has helped to localize specialized functions in the brain after brain damage, and it has gained new prominence through the arrival of modern optogenetic perturbation techniques that allow probing the functional contributions of neural circuit elements at unprecedented levels of detail. While inferences drawn from brain lesions are conceptually powerful, they face methodological difficulties due to the brains complexity. Particularly, they are challenged to disentangle the functional contributions of individual neural elements because many elements may contribute to a particular function, and these elements may be interacting anatomically as well as functionally. Therefore, studies of real-world data, as in clinical lesion studies, are not suitable for establishing the reliability of lesion approaches due to an unknown, potentially complex ground truth. Instead, ground truth studies of well-characterized artificial systems are required. Here, we systematically and exhaustively lesioned a small Artificial Neural Network (ANN) playing a classic arcade game. We determined the functional contributions of all nodes and links, contrasting results from single-element perturbations and perturbing multiple elements simultaneously. Moreover, we computed pairwise causal functional interactions between the network elements, and looked deeper into the systems inner workings, proposing a mechanistic explanation for the effects of lesions. We found that not every perturbation necessarily reveals causation, as lesioning elements, one at a time, produced biased results. By contrast, multi-site lesion analysis captured crucial details that were missed by single-site lesions. We conclude that even small and seemingly simple ANNs show surprising complexity that needs to be understood for deriving a causal picture of the system. In the context of rapidly evolving multivariate brain-mapping approaches and inference methods, we advocate using in-silico experiments and ground-truth models to verify fundamental assumptions, technical limitations, and the scope of possible interpretations of these methods. Author summaryThe motto "No causation without manipulation" is canonical to scientific endeavors. In particular, neuroscience seeks to find which brain elements are causally involved in cognition and behavior of interest by perturbing them. However, due to complex interactions among those elements, this goal has remained challenging. In this paper, we used an Artificial Neural Network as a ground-truth model to compare the inferential capacities of lesioning the system one element at a time against sampling from the set of all possible combinations of lesions. We argue for employing more exhaustive perturbation regimes since, as we show, lesioning one element at a time provides misleading results. We further advocate using simulated experiments and ground-truth models to verify the assumptions and limitations of brain-mapping methods.