Fantastic DNimals and where to find them

The introduction of deep neural networks (DNNs, Krizhevsky et al., 2012; Lecun et al., 1998) has altered the fields of computer vision and machine learning and these networks are starting to have an impact on the field of biological vision. Kendrick Kay's paper in this issue is therefore timely in addressing two important questions in this area, namely: i) how DNNs can be used to study vision, ii) how they relate to other modeling approaches. A central question Kay (2017) asks is how DNNs relate to the tri-level framework of Marr (1982). Within this framework a differentiation is made between the level of computational theory (what is the goal and logic of the process being carried out?), the representation and algorithm level (what is the representation of the input and output and how is this processed?), and the implementation level (how is this process implemented in hardware?). Kay asks to what degree it can be reasonably argued that DNNs are a good functional or mechanistic model of biological visual system. He finds them, beyond the most trivial inferences, lacking in this regard (Kay, 2017, section 8). I fully agree that it is problematic to relate DNNs to Marr's modeling levels, but at the same time I do not see this as a fundamental problem regarding the use of DNNs for understanding vision. Instead, I propose we consider DNNs in a similar way to how we consider animal models. An animal model for cognition typically emerges when an animal shows behavior that can be studied in a systematic fashion; such amodel increases in importance when the underlying anatomy and/or physiology of the animal canbe linked to thebehavior of interest.Animalmodels are relevant for understanding the human brain because they provide the opportunity to systematically measure and disrupt neural activity which is not possible with human subjects. Insights from such models can subsequently be applied in the development of different types of Marrian models. Implementations of neural network models are currently being studied because they have an impressive performance in identifying objects and show a wide range of ‘behavior’ similar to that of humans (Kheradpisheh et al., 2016; Kubilius et al., 2016). Methods for analysing these networks are being developed at a rapid rate (Kietzmann et al., 2017; Zintgraf et al., 2016) and are yielding insight into how these networks operate and how their behavior is related to architectural elements and computations. These results are used as measurements from a model organism, and do not require, nor suggest, that DNNs are implementation models of the animal visual system.