Getting ahead: forward models and their place in cognitive architecture

The use of forward models (mechanisms that predict the future state of a system) is well established in cognitive and computational neuroscience. We compare and contrast two recent, but interestingly divergent, accounts of the place of forward models in the human cognitive architecture. On the Auxiliary Forward Model (AFM) account, forward models are special-purpose prediction mechanisms implemented by additional circuitry distinct from core mechanisms of perception and action. On the Integral Forward Model (IFM) account, forward models lie at the heart of all forms of perception and action. We compare these neighbouring but importantly different visions and consider their implications for the cognitive sciences. We end by asking what kinds of empirical research might offer evidence favouring one or the other of these approaches. Two roles for forward models There is a great deal of evidence that people predict both themselves and other people [1,2]. But how do they do it? Recent proposals suggest that people use forward models [3–5] to make predictions (see Glossary). However, there are two different types of account of how they might be used. The first account (Figure 1) assumes a dedicated prediction mechanism implemented by additional circuitry distinct from the core mechanisms of perception and action. Those core mechanisms involve one or more distinct inverse models, which compute motor commands from desired effects. We call this complex the AFM account. The second account (Figure 2) is more integrated and posits a forward (generative) model as the core machinery of perception and action. We call this the IFM account. In the IFM, motor commands are replaced by the descending web of sensory predictions issued by the forward model. This removes any fundamental distinction between motor and sensory processing and sidesteps the need for a distinct inverse model or for the learning and use of multiple (paired) forward and inverse models. The two accounts thus share a great deal, but the differences between them are important, making them ripe (we argue) for directly contrastive research and experiment. AFMs When I plan an action, for instance moving my arm to a target, I construct an action command, use that command to perform the action, and experience the sensory (including proprioceptive) consequences of that action. If I repeatedly perform the action, I can learn from my mistakes (e.g., changing the plan slightly if my arm just misses the target). Over time, I can predict that if I instigate …