Mechanistic philosophy and the use of deep neural networks in neuroscience

Abstract

behavior

Deep Learning has revolutionized artificial intelligence (AI) over the past decade (LeCun et. al. 2015). This has led to the creation of ‘neurally inspired’ deep-neural-networks (DNNs). DNNs are claimed to be biologically realistic in the sense of incorporating key mechanistic or architectural features of the brain, such as a hierarchical structure. Interestingly, they are also said to exhibit similar behavioral capacities, as they are capable of performing ‘near human-level’ on a variety of behavioral tasks (Kriegeskorte 2015). These similarities have led researchers to propose DNNs as biologically realistic models of behavior and brain function (ibid). In this paper, I argue that there are at least two concerns in relating DNNs to the brain. First, I suggest that DNNs might not, as they stand, exhibit biologically realistic behavior. Secondly, I argue that there are key mechanistic dissimilarities between biological and artificial neural networks that impedes a so-called modelmechanism- mapping relationship. Thus there is a mismatch between the two systems at both the behavioral and implementational levels. The explanatory status of DNNs is accordingly called into question. To defend this position, I make recourse to the mechanistic philosophy of science, a framework within which to assess a model’s explanatory status (Craver 2007).