Differentiating between Bayesian parameter learning and structure learning based on behavioural and pupil measures.

Within predictive processing two kinds of learning can be distinguished: parameter learning and structure learning. In Bayesian parameter learning, parameters under a specific generative model are continuously being updated in light of new evidence. However, this learning mechanism cannot explain how new parameters are added to a model. Structure learning, unlike parameter learning, makes structural changes to a generative model by altering its causal connections or adding or removing parameters. Whilst these two types of learning have recently been formally differentiated, they have not been empirically distinguished. The aim of this research was to empirically differentiate between parameter learning and structure learning on the basis of how they affect pupil dilation. Participants took part in a within-subject computer-based learning experiment with two phases. In the first phase, participants had to learn the relationship between cues and target stimuli. In the second phase, they had to learn a conditional change in this relationship. Our results show that the learning dynamics were indeed qualitatively different between the two experimental phases, but in the opposite direction as we originally expected. Participants were learning more gradually in the second phase compared to the first phase. This might imply that participants built multiple models from scratch in the first phase (structure learning) before settling on one of these models. In the second phase, participants possibly just needed to update the probability distribution over the model parameters (parameter learning).

Brain-Inspired Hardware Solutions for Inference in Bayesian Networks.

The implementation of inference (i.e., computing posterior probabilities) in Bayesian networks using a conventional computing paradigm turns out to be inefficient in terms of energy, time, and space, due to the substantial resources required by floating-point operations. A departure from conventional computing systems to make use of the high parallelism of Bayesian inference has attracted recent attention, particularly in the hardware implementation of Bayesian networks. These efforts lead to several implementations ranging from digital circuits, mixed-signal circuits, to analog circuits by leveraging new emerging nonvolatile devices. Several stochastic computing architectures using Bayesian stochastic variables have been proposed, from FPGA-like architectures to brain-inspired architectures such as crossbar arrays. This comprehensive review paper discusses different hardware implementations of Bayesian networks considering different devices, circuits, and architectures, as well as a more futuristic overview to solve existing hardware implementation problems.

Young children integrate current observations, priors and agent information to predict others' actions.

From early on in life, children are able to use information from their environment to form predictions about events. For instance, they can use statistical information about a population to predict the sample drawn from that population and infer an agent's preferences from systematic violations of random sampling. We investigated whether and how young children infer an agent's sampling biases. Moreover, we examined whether pupil data of toddlers follow the predictions of a computational model based on the causal Bayesian network formalization of predictive processing. We formalized three hypotheses about how different explanatory variables (i.e., prior probabilities, current observations, and agent characteristics) are used to predict others' actions. We measured pupillary responses as a behavioral marker of 'prediction errors' (i.e., the perceived mismatch between what one's model of an agent predicts and what the agent actually does). Pupillary responses of 24-month-olds, but not 18-month-olds, showed that young children integrated information about current observations, priors and agents to make predictions about agents and their actions. These findings shed light on the mechanisms behind toddlers' inferences about agent-caused events. To our knowledge, this is the first study in which young children's pupillary responses are used as markers of prediction errors, which were qualitatively compared to the predictions by a computational model based on the causal Bayesian network formalization of predictive processing.

Processing of Prediction Errors in Mentalizing Areas.

When seeing people perform actions, we are able to quickly predict the action's outcomes. These predictions are not solely based on the observed actions themselves but utilize our prior knowledge of others. It has been suggested that observed outcomes that are not in line with these predictions result in prediction errors, which require additional processing to be integrated or updated. However, there is no consensus on whether this is indeed the case for the kind of high-level social-cognitive processes involved in action observation. In this fMRI study, we investigated whether observation of unexpected outcomes causes additional activation in line with the processing of prediction errors and, if so, whether this activation overlaps with activation in brain areas typically associated with social-cognitive processes. In the first part of the experiment, participants watched animated movies of two people playing a bowling game, one experienced and one novice player. In cases where the player's score was higher or lower than expected based on their skill level, there was increased BOLD activity in areas that were also activated during a theory of mind task that participants performed in the second part of the experiment. These findings are discussed in the light of different theoretical accounts of human social-cognitive processing.

Can infants' sense of agency be found in their behavior? Insights from babybot simulations of the mobile-paradigm.

The development of a sense of agency is essential for understanding the causal structure of the world. Previous studies have shown that infants tend to increase the frequency of an action when it is followed by an effect. This was shown, for instance, in the mobile-paradigm, in which infants were in control of moving an overhead mobile by means of a ribbon attached to one of their limbs. These findings have been interpreted as evidence for a sense of agency early in life, as infants were thought to have detected the causal action-movement relation. We argue that solely the increase in action frequency is insufficient as evidence for this claim. Computer simulations are used to demonstrate that systematic, limb-specific increase in movement frequency found in mobile-paradigm studies can be produced by an artificial agent (a 'babybot') implemented with a mechanism that does not represent cause-effect relations at all. Given that a sense of agency requires representing one's actions as the cause of the effect, a behavior that is reproduced with this non-representational babybot can be argued to be, in itself, insufficient as evidence for a sense of agency. However, a behavioral pattern that to date has received little attention in the context of sense of agency, namely an additional increase in movement frequency after the action-effect relation is discontinued, is not produced by the babybot. Future research could benefit from focusing on patterns whose production cannot be reproduced by our babybot as these may require the capacity for causal learning.

One wouldn't expect an expert bowler to hit only two pins: Hierarchical predictive processing of agent-caused events.

Evidence is accumulating that our brains process incoming information using top-down predictions. If lower level representations are correctly predicted by higher level representations, this enhances processing. However, if they are incorrectly predicted, additional processing is required at higher levels to "explain away" prediction errors. Here, we explored the potential nature of the models generating such predictions. More specifically, we investigated whether a predictive processing model with a hierarchical structure and causal relations between its levels is able to account for the processing of agent-caused events. In Experiment 1, participants watched animated movies of "experienced" and "novice" bowlers. The results are in line with the idea that prediction errors at a lower level of the hierarchy (i.e., the outcome of how many pins fell down) slow down reporting of information at a higher level (i.e., which agent was throwing the ball). Experiments 2 and 3 suggest that this effect is specific to situations in which the predictor is causally related to the outcome. Overall, the study supports the idea that a hierarchical predictive processing model can account for the processing of observed action outcomes and that the predictions involved are specific to cases where action outcomes can be predicted based on causal knowledge.

To be precise, the details don't matter: On predictive processing, precision, and level of detail of predictions.

Many theoretical and empirical contributions to the Predictive Processing account emphasize the important role of precision modulation of prediction errors. Recently it has been proposed that the causal models used in human predictive processing are best formally modeled by categorical probability distributions. Crucially, such distributions assume a well-defined, discrete state space. In this paper we explore the consequences of this formalization. In particular we argue that the level of detail of generative models and predictions modulates prediction error. We show that both increasing the level of detail of the generative models and decreasing the level of detail of the predictions can be suitable mechanisms for lowering prediction errors. Both increase precision, yet come at the price of lowering the amount of information that can be gained by correct predictions. Our theoretical result establishes a key open empirical question to address: How does the brain optimize the trade-off between high precision and information gain when making its predictions?

Beta- and gamma-band activity reflect predictive coding in the processing of causal events.

In daily life, complex events are perceived in a causal manner, suggesting that the brain relies on predictive processes to model them. Within predictive coding theory, oscillatory beta-band activity has been linked to top-down predictive signals and gamma-band activity to bottom-up prediction errors. However, neurocognitive evidence for predictive coding outside lower-level sensory areas is scarce. We used magnetoencephalography to investigate neural activity during probability-dependent action perception in three areas pivotal for causal inference, superior temporal sulcus, temporoparietal junction and medial prefrontal cortex, using bowling action animations. Within this network, Granger-causal connectivity in the beta-band was found to be strongest for backward top-down connections and gamma for feed-forward bottom-up connections. Moreover, beta-band power in TPJ increased parametrically with the predictability of the action kinematics-outcome sequences. Conversely, gamma-band power in TPJ and MPFC increased with prediction error. These findings suggest that the brain utilizes predictive-coding-like computations for higher-order cognition such as perception of causal events.

Free energy minimization and information gain: The devil is in the details.

Contrary to Friston's previous work, this paper describes free energy minimization using categorical probability distributions over discrete states. This alternative mathematical framework exposes a fundamental, yet unnoticed challenge for the free energy principle. When considering discrete state spaces one must specify their granularity, as the amount of information gain is defined over this state space. The more detailed this state space, the lower the precision of the predictions will be, and consequently, the higher the prediction errors. Hence, an optimal trade-off between precision and detail is needed, and we call for incorporating this aspect in the free energy principle.

Higher-level processes in the formation and application of associations during action understanding.

The associative account described in the target article provides a viable explanation for the origin of mirror neurons. We argue here that if mirror neurons develop purely by associative learning, then they cannot by themselves explain intentional action understanding. Higher-level processes seem to be involved in the formation of associations as well as in their application during action understanding.

Intentional communication: computationally easy or difficult?

Human intentional communication is marked by its flexibility and context sensitivity. Hypothesized brain mechanisms can provide convincing and complete explanations of the human capacity for intentional communication only insofar as they can match the computational power required for displaying that capacity. It is thus of importance for cognitive neuroscience to know how computationally complex intentional communication actually is. Though the subject of considerable debate, the computational complexity of communication remains so far unknown. In this paper we defend the position that the computational complexity of communication is not a constant, as some views of communication seem to hold, but rather a function of situational factors. We present a methodology for studying and characterizing the computational complexity of communication under different situational constraints. We illustrate our methodology for a model of the problems solved by receivers and senders during a communicative exchange. This approach opens the way to a principled identification of putative model parameters that control cognitive processes supporting intentional communication.