Prediction error minimization as a common computational principle for curiosity and creativity

Abstract

We propose expanding the authors’ shared novelty-seeking basis for creativity and curiosity by emphasizing an underlying computational principle: Minimizing prediction errors (mismatch between predictions and incoming data). Curiosity is tied to the anticipation of minimizing prediction errors through future, novel information, whereas creative AHA moments are connected to the actual minimization of prediction errors through current, novel information.

The authors Ivancovsky, Baror & Bar aim to reconcile the phenomena of creativity and curiosity via a shared novelty-seeking basis. For this, they describe common cognitive key features such as memory, cognitive control, attention and reward including empirical evidence from network neuroscience. We agree with their efforts to reconcile those concepts under one explanatory framework. However, we argue that it can be further expanded by linking the novelty-seeking basis to the process of prediction error minimization – a common underlying computational principle both central to predictive coding (Clark, 2013) and reinforcement learning theories (Glimcher, 2011). This principle connects both phenomena as has been argued before (Friston et al., 2017; Van de Cruys et al., 2021).

According to predictive coding theories, our sensory and cognitive systems have a fundamental aim: Constructing reliable representations of the world to enable adaptive behaviour. To this end, the brain engages in a process of generating predictions about its own perceptual experiences, actions and cognitive processes serving as models of those experiences and processes (Den Ouden, Kok, & De Lange, 2012). These predictions are subsequently compared with the corresponding incoming sensory, motor or cognitive input, resulting in the computation of a prediction error (Den Ouden et al., 2012). The bigger the mismatch between input and predictions, the bigger the resulting error, prompting an update to those predictive models which ultimately forms the basis for learning (Friston & Kiebel, 2009). Prediction errors can be unsigned, representing the magnitude of the surprise related to a perception or cognitive outcome. Prediction errors can also be signed, indicating the valence of the outcome (whether it is better or worse than expected) often related to reward (Den Ouden et al., 2012). Importantly, the concept of prediction errors are inherently related to novelty because minimizing prediction errors by updating one's models entails constantly seeking out unexpected, novel information. This process again leads to more prediction errors generating a continuous cycle of learning and adaptation (Schwartenbeck, FitzGerald, Mathys, Dolan, & Friston, 2015).

From this perspective, curiosity and creative problem solving are two phenomena that relate to different aspects within this same continuous prediction error minimization process. Consistent with this view, the phenomenology and neurobiology of curiosity (Gruber, Gelman, & Ranganath, 2014; Gruber & Ranganath, 2019) and creative problem solving (Becker, Wang, & Cabeza, 2023; Dubey, Ho, Mehta, & Griffiths, 2021; Friston et al., 2017; Savinova & Korovkin, 2022), have both been explained via different kinds of signed and unsigned prediction error signals. In the following, we argue that curiosity reflects expected information gain while creative problem solving or at least its end result – the AHA experience – represents absolute information gain (Van de Cruys et al., 2021). Information gain quantifies how much a model is updated due to new information causing a prediction error.

Researchers commonly define curiosity as a motivational state that stimulates exploration and information seeking to reduce uncertainty (Gruber & Ranganath, 2019; Ivancovsky, Baror, & Bar). When we encounter something unexpected, like a clown at a professional gathering or an unfamiliar problem, it can trigger our curiosity and motivate us to explore novel information, possibly to understand the clown's presence or to attempt to solve the problem (Friston et al., 2017). It has been suggested that curiosity is triggered by strong prediction errors that are seen as indicators of potentially valuable future information (Gruber & Ranganath, 2019). Essentially, curiosity can be characterized as expected information gain, where prediction errors arising from unexpected events, such as encountering an unfamiliar problem, provide an estimate of how much a new piece of information (its solution) is expected to minimize these prediction errors, leading to a model update. Note, the more substantial the expected model update, the higher the new information's expected gain.

Creativity involves breaking away from typical expectations and generating novel and useful ideas or solutions (Mednick, 1962). Insight is a fundamental process in creative problem-solving that occurs when a non-obvious problem is solved via a novel solution approach often eliciting an AHA experience (Danek, Williams, & Wiley, 2020; Dietrich & Kanso, 2010). The AHA experience describes the solver's conviction that the solution arrived suddenly, is surprising, certainly correct, involves a feeling of pleasure and internal reward (Kizilirmak & Becker, 2023). Due to its close conceptual proximity to surprise and reward, the AHA! experience has recently been reframed as a combination of different prediction errors tied to different aspects of the problem-solving process, such as the timing of a solution, the solvability of the problem or accuracy of the solution content (Becker et al., 2023; Dubey et al., 2021; Friston et al., 2017). For example, it is assumed that individuals maintain a metacognitive model of their abilities that predicts when they will solve a problem. A prediction error occurs when the solution is found faster than expected, generating a sense of surprise and internal reward (Dubey et al., 2021). In that sense, the AHA experience during creative problem solving reflects the actual information gain. Actual information gain here describes the (not expected but) actual size of the prediction errors caused by the new piece of information (the solution) leading to a model update and ultimately to a more reliable representation of the world.

In sum, we argued that the shared novelty-seeking basis of curiosity and creativity can be related to one underlying computational principle of prediction error minimization. Curiosity corresponds to an expected gain (model update) for new information that has not yet emerged but whose size is estimated by a current prediction error. In contrast, the AHA experience in creative problem-solving corresponds to the actual gain (model update) of new information that has just become available resulting in a prediction error. In reinforcement learning and the predictive coding theories, this principle has been associated with several other phenomena, such as perception, decision making under uncertainty, memory and reversal, habit or reward-based learning (Friston et al., 2017; Glimcher, 2011). Considering Ivancovsky et al.'s effort to reconcile creativity and curiosity through a common novelty-seeking basis, we believe that this computational principle represents an important perspective for further consideration.