AAP 2026 CONFERENCE

Causation in classical computation is relatively simple – it involves a series of events, each causally dependent on its predecessor. Causation in neural networks can be more complex. One reason for this is the possibility of recurrent or re-entrant signals. This paper investigates this topic. First, drawing on the analysis of Anne Treisman, I look at the role that recurrence may play in binding – which in this context can be thought of as the combination of simple representations into more complex representations. Second, I analyse the kinds of causal patterns involved here through Luigi Pasinetti’s work, recently translated into English, on causal dependence and interdependence. The idea, roughly speaking, is that neural networks that bind simple representations into complex representations through some kind of recurrent activity will be comprised of states in relationships of interdependence, rather than the relations of casual dependence that characterise classical computation. In this way, it is suggested, Pasinetti’s concepts can be used to distinguish the causal patterns that characterise classical computation from those that characterise the kinds of neural networks described here.