1. Introduction 

1.1. Intentional inference: perceptual information and top-down prior knowledge

Explaining or predicting the behaviour of our conspecifics requires the ability to properly appreciate the causes that motivate it. As the observed behaviour is self-initiated and directed at a specific goal, these causes are hidden rather than visible: intentions – like beliefs or desires – are unobservable states. Thus, it has long been a matter of speculation how one may infer them from mere observation, that is, from patterns of visible behaviour alone.

One decade ago, this question got a resounding response with the discovery, in the monkey and then in humans, of a system of cerebral regions endowed with ‘mirror’ properties. Neurons from these regions discharge both when a subject performs a given action and sees the same action performed by a third party (Gallese, Fadiga, Fogassi, & Rizzolatti, 1996; Rizzolatti, Fadiga, Gallese, & Fogassi, 1996; Grafton, Arbib, Fadiga, & Rizzolatti, 1996). This process of "motor equivalence" speaks convincingly in favor of a direct relationship between action observation and action understanding. During action observation, visual information conveyed by movement kinematics automatically maps onto the observer’s own motor system, then activating representations similar to those the subject would have generated during active action (Rizzolatti & Craighero, 2004). The activation of these ‘shared’ representations – activation that is similar to a process of internal replication, known as ‘simulation’ – would provide a direct access to the intentions of the agent, thus allowing prediction to be made about her current behaviour (Gallese & Goldman, 1998; Grèzes & Decéty, 2001;Rizolatti, Fogassi, & Gallese, 2001;Jeannerod, 2001;Blakemore & Decéty, 2001).

However, a problem arises with this model. In the “mirror” account of action understanding, visual information from the observed movement is supposed to be ‘directly’ transformed into knowledge about the action goal by inverting the model we generally use to perform an action (from current visual kinematics to action plans congruent with this kinematics). Thus, a one-to-one mapping is implicitly assumed between intention and action: one particular action denotes one particular intention that can be inferred once the corresponding visual kinematics is properly processed. But as a rule, the relation between mental states and observable behaviour is more complex than a one-to-one correspondence (Oztop, Wolpert, & Kawato, 2005; Jacob, 2008; see also Becchio, Sartori, Bulgheroni, & Castiello, 2008). Besides, information conveyed by kinematic movement is often noisy, ambiguous, or incomplete. As a result, visual information generally under-constraints the space of candidate causes (i.e. the many competing intentions) that are logically consistent with what is observed (Kilner, Friston, & Frith, 2007a, 2007b; Csibra & Gergely, 2007).

One way to solve this problem is to assume that this space of possible intentions is further constrained by the observer’s prior expectations. These expectations are derived from a prior knowledge that may originate from the past experience of the viewer (through expertise: Sebanz & Shiffrar, 2006; Aglioti, Cesari, Romani, & Urgesi, 2008, or learning of statistical regularities: Gold & Shadlen, 2002), from her intuitive theories (Carey, 1985; Gopnik & Wellman, 1994), or reputational knowledge (King-Casas et al., 2005; Singer et al., 2006), as well as from contextual information surrounding the action scene (Sebanz et al., 2006; Aglioti et al., 2008). This prior knowledge has been demonstrated crucial to account for the robustness of our everyday inferences (see Frith & Frith, 2006). Indeed, by exerting top-down influences onto the observer’s shared representations, it makes possible inductive inference about the agent’s intentions, even in case of noisy signals or incomplete data(Baker, Tenenbaum, & Saxe, 2006, 2007; Griffiths, Kemp, & Tenenbaum, 2008).

However, although most authors agree that prior knowledge and perceptual information both contribute to the process of inferring intentions, the precise contribution of each type of information remains controversial (see Gallese, Keysers, & Rizzolatti, 2004; Jacob & Jeannerod, 2005; Saxe, 2005; Goldman & Sebanz, 2005; Mitchell, 2005; Nichols & Stich; Keysers & Gazzola, 2007; deLange, Spronk, Willems, Toni, & Bekkering, 2008). The controversy appears to stem in part from the fact that "intention" is an umbrella term used in the empirical and philosophical literature to refer to representations of actions that can differ in both their content and format, as well as in their temporal properties and in the role they play in the guidance of actions (Searle, 1983; Bratman, 1987; Mele, 1992; Pacherie 2000, 2008; Jeannerod & Pacherie, 2004). Intentions can therefore be distinguished into several sub-types according to one or several of these factors. In the present study, we propose a typology of intentions and hypothesize that this typology might be a key element in understanding how perceptual information and prior knowledge contribute to the process by which intentions are inferred.