Probabilistic and predictive models

Our work on internal predictive models€—neural systems that simulate the behavior of a natural process—has shown that they are a vital theoretical component used to solve fundamental problems in sensorimotor control.

We have shown that the human motor system uses efference copy of the motor commands (i.e., a copy of the outgoing motor command) to construct an internal forward model, which is a key processing component in human state estimation (Wolpert et al., 1995). In addition, we showed in a series of studies that prediction is used to filter incoming sensory information, removing predictable components, thereby enhancing the detection of novel stimuli and that this predictive process involves cerebellar circuitry (Blakemore et al., 1998; Blakemore et al., 1999; Blakemore et al., 2001). We have also developed a model in which impairments in the internal representations of actual, desired and predicted states can explain a broad range of neuropsychiatric symptoms (Frith et al., 2000). To examine this model in patient populations, we developed a new paradigm in which sensory cancellation can be assessed in individuals (Shergill et al., 2003) and, as hypothesized, demonstrated deficits in prediction in this patient group (Shergill et al., 2005).

We have extended the idea of internal models within the framework of Bayesian Decision Theory, which is a probabilistic framework in which prior knowledge is optimally combined with sensory evidence to generate beliefs about the world. We have shown that the sensorimotor system uses such a probabilistic approach to handle information in an attempt to behave optimally in the face of uncertainty (Kording and Wolpert, 2004; Faisal et al., 2008; Franklin and Wolpert, 2011; Orban and Wolpert, 2011). Recently, we have developed a novel method, termed cognitive tomography, to extract high dimensional priors from low dimensional responses which show that priors can be highly subject-specific yet, importantly, generalize across tasks (Houlsby et al., 2013). We have demonstrated that sensorimotor predictions are affected by aging (Wolpe et al., 2016) and that deficits in prediction may underpin symptoms of psychiatric disease (Teufel et al., 2010; Shergill et al., 2013; Shergill et al., 2014). We have also identified situations in which healthy humans act sub-optimally, therefore exposing limitations in probabilistic processing (Acerbi et al., 2012; Acerbi et al., 2014).