Ideal Patterns and Non-Factive Understanding

This paper begins with the observation of a Levins style trade-off in models of complex phenomena, between predictive accuracy of the model, on the one hand, and intelligibility (the capacity of the model to provide understanding of the phenomenon under investigation), on the other. I provide examples from recent use of connectionist models in neuroscience, which are predictively very accurate but less intelligible than earlier generations of models. The existence of this trade-off lends support to non-factivist accounts of scientific understanding. However, non-factivism faces important objections from Khalifa (2017) and Sullivan and Khalifa (forthcoming). I reply to these objections by arguing that a critical weakness in the non-factivist account of Potochnik (2017) comes from the use of Dennett’s (1991) notion of a “real pattern.” I show that the account can be strengthened by replacing this with the notion of an “ideal pattern”, where the phenomena that are the targets of model building do not comprise patterns that are simply “out there” in nature, but are to some extent dependent on the methods of data-processing chosen by the researcher.