Linking the brain with behaviour: the neural dynamics of success and failure in goal-directed behaviour

Amanda K. Robinson, Anina N. Rich, Alexandra WoolgarbioRxiv 2021 · DOI

TL;DR

This study uses MEG and advanced decoding techniques to show how the brain processes information during a complex task, revealing a clear temporal progression from sensory input to decision-making. Crucially, it demonstrates that when errors occur, neural representations can shift from reflecting the actual stimulus to representing the incorrect stimulus, providing a direct link between neural activity and behavioral outcomes.

What Problem Does This Paper Address?

Understanding how the brain makes decisions and controls behavior is a central challenge in neuroscience. While we know that areas like the frontoparietal cortex are involved in cognitive control, it's less clear how information is processed over time, especially when things go wrong. Specifically, researchers want to know if the neural patterns we observe are truly meaningful for behavior, or if they're just byproducts. This paper tackles this by investigating the temporal dynamics of information coding in the brain during a challenging task and examining how these neural patterns change when participants make errors, aiming to establish a direct link between neural activity and behavioral success or failure.

Methods

The researchers employed magnetoencephalography (MEG) to record brain activity with high temporal resolution. They then used time-resolved multivariate pattern analysis (MVPA), a powerful decoding technique. MVPA works by training a classifier to distinguish between different neural patterns associated with different stimuli or tasks. By applying this to MEG data across many time points, they could track when specific information (like stimulus location, cue color, or task rule) was represented in the brain.

A key aspect of their approach was response-aligned analysis. Unlike typical analyses that align to stimulus onset, aligning to the response time helps capture processes that are more directly related to the final decision and motor output, especially in tasks with variable reaction times. They also analyzed error trials specifically, training classifiers on correct trials and testing on incorrect ones to see how neural representations differed when mistakes were made.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Neural representations can "reverse" during errors: This is perhaps the most striking finding. When participants made a stimulus-related error, the neural activity initially reflected the presented stimulus but then shifted to represent the incorrect stimulus, the one the participant chose. This suggests the brain actively constructs a representation that leads to the error, rather than just misinterpreting the input.
  • Information processing is a temporal cascade: The study clearly shows a progression of information coding. Sensory information (stimulus position) appears early, followed by rule information, and then motor preparation, all unfolding over time.
  • Response-aligned analysis is crucial for understanding decisions: Aligning neural data to the response time, rather than just stimulus onset, revealed processes like evidence accumulation and motor preparation that are critical for decision-making but can be masked by stimulus-onset timing variability.
  • Rule coding is less robust than stimulus coding, especially during errors: While stimulus information was strongly represented and even reversed during errors, rule information was more transient and less evident, particularly when mistakes were made.

Key Results

  • Correct Trials: Stimulus position information was decoded as early as 75ms after stimulus onset. Cue information was decoded from 170ms. Rule information was evident around 150ms and more prominently from 400-1200ms. Response information was decoded around 715ms.
  • Error Trials (Stimulus Errors):
    • Onset-aligned: Initial stimulus coding reflected the correct stimulus, but later coding (after 335ms) became transient and then, after 495ms, showed substantial evidence for coding the incorrect stimulus.
  • Response-aligned: Neural patterns ramped towards the incorrect stimulus representation from approximately 355ms before the response.
  • Error Trials (Rule Errors): Stimulus coding remained similar to correct trials, but rule coding was weak and not sustained, showing little evidence of reversal.

Limitations and Future Work

The authors acknowledge that rule coding was generally weak and difficult to interpret with confidence, especially on error trials. This might be due to the small number of rule error trials, the inherent variability in timing for rule-based tasks, or limitations in MEG signal quality from frontal regions.

Future work could explore the spatial extent of these processes more precisely, potentially by combining fMRI and MEG data. Investigating how these dynamics change in other complex cognitive tasks and with different types of errors would also be valuable. The study also highlights the need for more sophisticated computational methods to analyze the complex, overlapping dynamics of information coding in the brain.

Summary updated

Abstract

Abstract The human brain is extremely flexible and capable of rapidly selecting relevant information in accordance with task goals. Regions of frontoparietal cortex flexibly represent relevant task information such as task rules and stimulus features when participants perform tasks successfully, but less is known about how information processing breaks down when participants make mistakes. This is important for understanding whether and when information coding recorded with neuroimaging is directly meaningful for behaviour. Here, we used magnetoencephalography (MEG) to assess the temporal dynamics of information processing, and linked neural responses with goal-directed behaviour by analysing how they changed on behavioural error. Participants performed a difficult stimulus-response task using two stimulus-response mapping rules. We used time-resolved multivariate pattern analysis to characterise the progression of information coding from perceptual information about the stimulus, cue and rule coding, and finally, motor response. Response-aligned analyses revealed a ramping up of perceptual information prior to a correct response, suggestive of internal evidence accumulation. Strikingly, when participants made a stimulus-related error, and not when they made other types of errors, patterns of activity initially reflected the stimulus presented, but later reversed, and accumulated towards a representation of the incorrect stimulus. This suggests that the patterns recorded at later timepoints reflect an internally generated stimulus representation that was used to make the (incorrect) decision. These results illustrate the orderly and overlapping temporal dynamics of information coding in perceptual decision-making and show a clear link between neural patterns in the late stages of processing and behaviour.
BibTeX
@misc{robinson2021linking,
  author = {Amanda K. Robinson and Anina N. Rich and Alexandra Woolgar},
  title = {{Linking the brain with behaviour: the neural dynamics of success and failure in goal-directed behaviour}},
  year = {2021},
  doi = {10.1101/2021.05.25.445701},
  howpublished = {bioRxiv},
}
Figure