Frederik Junker
Affiliation: University of Copenhagen
Category: Philosophy
Keywords: Inference, Deliberation, Representational formats, Mental simulation, Cognitive maps, Metacognition, Rationality
Date: Thursday 4th of September
Time: 15:00
Location: Gen. Henryk Dąbrowski Hall (006)
View the full session: Spatial Cognition
We infer new beliefs from existing ones and decide on actions based on what we believe and desire. But what is required to draw an inference? In paradigm cases, we are aware of the premises on which we base our conclusion and of some formal rule (e.g., modus ponens) that governs the transition from premises to conclusion (Boghossian, 2014). However, it is becoming increasingly clear that inferences can be drawn under less stringent conditions, prompting the question: what are the lower bounds of inference? The conditions for inference can be relaxed along at least two dimensions: awareness and representational format. Some argue that inference can occur even when we are unaware of the premise states or unaware that we are drawing an inference (Levy, 2024; Quilty-Dunn & Mandelbaum, 2018; Siegel, 2019). This perspective allows for inference without conscious deliberation. Others contend that inference can occur without the transitions being governed by formal rules (Buckner, 2019; Shea, 2023). This view accommodates inferential transitions over representations with non-discursive formats that lack logical structure.
In this paper, I argue for relaxing the bounds on inference, a move that raises a number of further questions. Expanding the scope of inference suggests that it is more pervasive than typically assumed in paradigm cases. If inference extends more broadly, it seems that the frequency with which we form and revise beliefs and intentions – key outputs of inference – would also increase. However, this raises concerns about the viability of long-term planning. If we were prone to frequently canceling our intentions or lowering our confidence in success, we might struggle to achieve long-term goals that require stable commitment over time (Bratman, 1987). Additionally, relaxing the bounds on inference brings into question the issue of warrant: if inference can occur without awareness or adherence to rules, how can we be confident that our inferences are – or generally tend to be – warranted?
To motivate a broader conception of inference, I examine some less demanding candidates. First, I consider transitions between mental images during mental simulation, which enable us to derive new consequences from stored representations. Second, I discuss the format and function of cognitive maps. Cognitive maps are structural representations that encode relations between spatial properties (e.g., locations) or more abstract, non-spatial properties (e.g., the competence and popularity of prospective partners). These representations facilitate spatial planning and generalizations of abstract relations to new situations (Behrens et al., 2018).
Both mental simulation and map-based transitions can yield the standard outputs of inference: new or revised beliefs and intentions. Neither process requires awareness of these transitions, nor do they depend on discursive formats or rule-governed transitions. On more permissive accounts, this suffices for inference. One concern, however, is that these mechanisms make attitude changes so pervasive that they threaten long-term planning. Another is whether simulation- or map-based inferences come with any warrant. I argue that both concerns can be addressed by appealing to certain constraints on these forms of inference.
Firstly, certain kinds of constraints are built into the representations employed in mental simulations and cognitive maps. Mental simulations are constrained by mental models that approximate the causal structure of the world (Battaglia et al., 2013; Ullman et al., 2017). Similarly, cognitive maps encode structural features of the world, such as spatial relationships and distances between locations. These world-reflecting structures contribute to the reliability of inferences drawn from mental simulations and cognitive maps.
Secondly, mental simulations and cognitive maps can generate reasons for and against certain conclusions which constrain subsequent inferences. Both mental simulations (Bulley & Schacter, 2020) and cognitive maps (Liu et al., 2021) play crucial roles in evaluating potential actions and outcomes and in identifying which course of actions has the highest expected value – determined by the likelihood of success, the value of a successful outcome, and the associated costs. Learning that an action has higher expected value than the alternatives provides the agent with a reason to pursue it. Additionally, mental simulations can make otherwise inaccessible information accessible to higher cognitive processes (Aronowitz & Lombrozo, 2020; Miyazono & Tooming, 2023) and flexibly recombine stored representations to form representations of novel scenarios (Shea, 2022). Mental simulation and cognitive maps thereby help the agent learn both how certain conclusions can be reached and in virtue of what they are justified.
Finally, metacognitive feelings help constrain inferences, such that conclusions are most likely to be drawn when accompanied by appropriate metacognitive feelings. For instance, feelings of reliability can help restrict inferences to cases where they are accompanied by feelings of high reliability. By calibrating these feelings according to their their downstream consequences – strengthening after successful inferences and weakening after unsuccessful ones – they come to track actual reliability, thereby ensuring that the inferences are warranted. Since metacognitive feelings also constrain unconscious, non-logical inferences (Shea, 2023), they likely play a similar role in simulation- and map-based inferences.
These constraints explain why simulation- and map-based inferences do not undermine the stability of beliefs and intentions needed to reach long-term goals. These inference types track worldly structures and are responsive to reasons. Since sticking to long-term goals often has high expected value, and reevaluations only give us reason to abandon them when superior alternatives are discovered, our reasons for maintaining these goals typically persist. Moreover, metacognitive feelings ensure that not just any simulation- or map-based transition leads to an inference. Our cognitive processes are monitored and regulated so that we are most inclined to draw a conclusion when the transition is accompanied by a feeling of reliability, while those deemed unreliable are frequently blocked.
In summary, the role of mental simulation and cognitive maps in reaching conclusions supports a broader conception of inference, recognizing these processes as genuinely inferential. Concerns that this expansion would allow an influx of unwarranted inferences are misplaced, as the cognitive constraints governing these processes ensure that they remain rational and properly bounded.