Graph-Attention-Based Casual Discovery With Trust Region-Navigated Clipping Policy Optimization

In many domains of empirical sciences, discovering the causal structure within variables remains an indispensable task. Recently, to tackle unoriented edges or latent assumptions violation suffered by conventional methods, researchers formulated a reinforcement learning (RL) procedure for causal discovery and equipped a REINFORCE algorithm to search for the best rewarded directed acyclic graph. The two keys to the overall performance of the procedure are the robustness of RL methods and the efficient encoding of variables. However, on the one hand, REINFORCE is prone to local convergence and unstable performance during training. Neither trust region policy optimization, being computationally expensive, nor proximal policy optimization (PPO), suffering from aggregate constraint deviation, is a decent alternative for combinatory optimization problems with considerable individual subactions. We propose a trust region-navigated clipping policy optimization method for causal discovery that guarantees both better search efficiency and steadiness in policy optimization, in comparison with REINFORCE, PPO, and our prioritized sampling-guided REINFORCE implementation. On the other hand, to boost the efficient encoding of variables, we propose a refined graph attention encoder called SDGAT that can grasp more feature information without priori neighborhood information. With these improvements, the proposed method outperforms the former RL method in both synthetic and benchmark datasets in terms of output results and optimization robustness.

The Temporal Order of Word Presentation Modulates the Amplitudes of P2 and N400 during Recognition of Causal Relations

The processing of causal relations has been constantly found to be asymmetrical once the roles of cause and effect are assigned to objects in interactions. We used a relationship recognition paradigm and recorded electroencephalographic (EEG) signals to explore the neural mechanism underlying the asymmetrical representations of causal relations in semantic memory. The results revealed that the verification of causal relations is faster if two words appear in “cause-effect” order (e.g., virus-epidemic) than if they appear in “effect-cause” order (e.g., epidemic-virus), whereas no such asymmetrical representation was found for the verification of hierarchical relations with reverse orders (e.g., bird-sparrow vs. sparrow-bird) in Experiment 1. Furthermore, the P2 amplitude elicited by “superordinate-subordinate” order was larger than that when in reverse order, whereas the N400 effect elicited by “cause-effect” order was smaller (more positive) than when in reverse order. However, no such asymmetry, as well as P2 and N400 components, were observed when verifying the existence of a general associative relation in Experiment 2. We suggested that the smaller N400 in cause-effect order indicates their increased salience in semantic memory relative to the effect-cause order. These results provide evidence for dissociable neural processes, which are related to role binding, contributing to the generation of causal asymmetry.

The Processing of Causal and Hierarchical Relations in Semantic Memory as Revealed by N400 and Frontal Negativity

Most current studies investigating semantic memory have focused on associative (ring-emerald) or taxonomic relations (bird-sparrow). Little is known about the question of how causal relations (virus-epidemic) are stored and accessed in semantic memory. The goal of this study was to examine the processing of causally related, general associatively related and hierarchically related word pairs when participants were required to evaluate whether pairs of words were related in any way. The ERP data showed that the N400 amplitude (200-500 ms) elicited by unrelated related words was more negative than all related words. Furthermore, the late frontal distributed negativity (500-700 ms) elicited by causally related words was smaller than hierarchically related words, but not for general associated words. These results suggested the processing of causal relations and hierarchical relations in semantic memory recruited different degrees of cognitive resources, especially for role binding.

Electrophysiological difference between the representations of causal judgment and associative judgment in semantic memory

Causally related concepts like "virus" and "epidemic" and general associatively related concepts like "ring" and "emerald" are represented and accessed separately. The Evoked Response Potential (ERP) procedure was used to examine the representations of causal judgment and associative judgment in semantic memory. Participants were required to remember a task cue (causal or associative) presented at the beginning of each trial, and assess whether the relationship between subsequently presented words matched the initial task cue. The ERP data showed that an N400 effect (250-450 ms) was more negative for unrelated words than for all related words. Furthermore, the N400 effect elicited by causal relations was more positive than for associative relations in causal cue condition, whereas no significant difference was found in the associative cue condition. The centrally distributed late ERP component (650-750 ms) elicited by the causal cue condition was more positive than for the associative cue condition. These results suggested that the processing of causal judgment and associative judgment in semantic memory recruited different degrees of attentional and executive resources.

Neural correlates of causal power judgments

Causal inference is a fundamental component of cognition and perception. Probabilistic theories of causal judgment (most notably causal Bayes networks) derive causal judgments using metrics that integrate contingency information. But human estimates typically diverge from these normative predictions. This is because human causal power judgments are typically strongly influenced by beliefs concerning underlying causal mechanisms, and because of the way knowledge is retrieved from human memory during the judgment process. Neuroimaging studies indicate that the brain distinguishes causal events from mere covariation, and also distinguishes between perceived and inferred causality. Areas involved in error prediction are also activated, implying automatic activation of possible exception cases during causal decision-making.

Contradictory reasoning network: an EEG and FMRI study

Contradiction is a cornerstone of human rationality, essential for everyday life and communication. We investigated electroencephalographic (EEG) and functional magnetic resonance imaging (fMRI) in separate recording sessions during contradictory judgments, using a logical structure based on categorical propositions of the Aristotelian Square of Opposition (ASoO). The use of ASoO propositions, while controlling for potential linguistic or semantic confounds, enabled us to observe the spatial temporal unfolding of this contradictory reasoning. The processing started with the inversion of the logical operators corresponding to right middle frontal gyrus (rMFG-BA11) activation, followed by identification of contradictory statement associated with in the right inferior frontal gyrus (rIFG-BA47) activation. Right medial frontal gyrus (rMeFG, BA10) and anterior cingulate cortex (ACC, BA32) contributed to the later stages of process. We observed a correlation between the delayed latency of rBA11 response and the reaction time delay during inductive vs. deductive reasoning. This supports the notion that rBA11 is crucial for manipulating the logical operators. Slower processing time and stronger brain responses for inductive logic suggested that examples are easier to process than general principles and are more likely to simplify communication.