Information theoretic evidence for predictive coding in the face processing system

Alla Brodski‐Guerniero, Georg-Friedrich Paasch, Patricia Wollstadt, İpek Özdemir, Joseph T. Lizier, Michael WibralbioRxiv 2016 · DOI

TL;DR

This study uses information theory to investigate how the brain uses prior knowledge to make predictions, specifically in face processing. They found that activated prior knowledge for faces is represented by predictable neural signals in face-specific brain areas, and this predictability is linked to faster reaction times.

What Problem Does This Paper Address?

Predictive coding theory suggests our brains constantly generate predictions about incoming sensory information based on prior knowledge. While this theory is influential, direct neurophysiological evidence for how this "prior knowledge" is maintained and used for predictions has been scarce. A key challenge is that traditional methods often require strong, pre-defined assumptions about which brain areas are involved and what specific predictions are being made. This paper aims to provide a more direct, assumption-free look at the neural signatures of activated prior knowledge and its role in predictive processing, focusing on face recognition.

Methods

The researchers employed information-theoretic measures, specifically Active Information Storage (AIS) and Transfer Entropy (TE), applied to Magnetoencephalography (MEG) data.

  • Active Information Storage (AIS): This measures how much information in a neural signal is predictable from its own past. The idea is that if prior knowledge is activated and maintained, the neural signals representing it should be more predictable over time.
  • Transfer Entropy (TE): This quantifies directed information flow between brain regions. It's used here to investigate how predictions are communicated, particularly in a top-down manner.

These measures were applied to whole-brain source-reconstructed MEG data. They also used Support Vector Machines (SVMs) for decoding analysis to see if the content of predictions (face vs. house) could be identified from AIS values. Finally, spectral Granger causality analysis was used to examine the frequency-specific nature of information transfer.

Datasets

The study used MEG data from 52 human subjects performing a face/house detection task. No new datasets were introduced.

Key Insights and Ideas

  • Predictable Neural Signals = Activated Prior Knowledge: The core idea is that activated prior knowledge, when maintained by the brain, should manifest as predictable patterns in neural activity. This predictability is what AIS measures.
  • Content-Specific Brain Regions: The study found that this "predictable information" was specifically elevated in brain areas known for face processing (like the fusiform face area - FFA) when subjects were engaged in a face detection task. This suggests that the brain doesn't just maintain general predictions, but content-specific ones.
  • Low-Frequency Activity is Key: The findings strongly support the idea that activated prior knowledge and predictions are primarily signaled in low-frequency brain activity (alpha and beta bands, <30 Hz).
  • Top-Down Prediction Transfer: Transfer entropy revealed increased information flow from higher-level areas (like PPC and alT) to lower-level face processing areas (FFA) during face prediction, suggesting a top-down communication of predictions. This transfer was particularly associated with beta frequencies.
  • Behavioral Relevance: The degree of predictable information in FFA was not only content-specific but also correlated with faster reaction times, indicating that this neural predictability directly facilitates performance.

Key Results

  • Increased AIS in Face Areas: During face blocks, AIS was significantly higher in the fusiform face area (FFA), occipital face area (OFA), anterior inferior temporal cortex (aIT), posterior parietal cortex (PPC), and primary visual cortex (V1) compared to house blocks. No such increase was found for house blocks.
  • Decoding Face Predictions: Support Vector Machines could decode the cued category (face or house) from trial-by-trial AIS values in the face prediction network with above-chance accuracy (median 65.2%, significantly better than chance for 22 out of 52 subjects after correction).
  • Alpha/Beta Band Correlation: A strong positive correlation was found between single-trial spectral power in the alpha and beta frequency bands and single-trial AIS, suggesting these frequencies are the likely carriers of activated prior knowledge.
  • Top-Down Information Transfer: Increased information transfer was observed from aIT to FFA and from PPC to FFA during face blocks, primarily in the beta frequency band. For house blocks, a bottom-up link from V1 to PPC showed increased transfer.
  • Behavioral Facilitation: In FFA, higher AIS (predictable information) during face blocks was significantly correlated with shorter reaction times (rho = -0.41), indicating that pre-activated face knowledge speeds up detection.

Limitations and Future Work

  • House Prediction Network: The study did not find a clear "house prediction network" with increased AIS, unlike the face network. The authors suggest this might be due to faces having more regularities for top-down processing compared to natural scenes.
  • Speculation on Bottom-Up House Transfer: The observed bottom-up link for house blocks (V1 to PPC) is speculative, as the analysis was primarily focused on the face prediction network.
  • Future Directions: The findings support microcircuit models of predictive coding, suggesting that predictions are transferred top-down via low-frequency channels and prediction errors bottom-up via high-frequency channels. Future work could further explore these spectral dissociations and the role of different frequency bands in predictive coding. The study also highlights the utility of information-theoretic measures for investigating brain function without strong prior assumptions.

Summary updated

Abstract

Abstract Predictive coding suggests that the brain infers the causes of its sensations by combining sensory evidence with internal predictions based on available prior knowledge. However, the neurophysiological correlates of (pre-)activated prior knowledge serving these predictions are still unknown. Based on the idea that such pre-activated prior knowledge must be maintained until needed we measured the amount of maintained information in neural signals via the active information storage (AIS) measure. AIS was calculated on whole-brain beamformer-reconstructed source time-courses from magnetoencephalography (MEG) recordings of 52 human subjects during the baseline of a Mooney face/house detection task. Pre-activation of prior knowledge for faces showed as alpha- and beta-band related AIS increases in content specific areas; these AIS increases were behaviourally relevant in brain area FFA. Further, AIS allowed decoding of the cued category on a trial-by-trial basis. Moreover, top-down transfer of predictions estimated by transfer entropy was associated with beta frequencies. Our results support accounts that activated prior knowledge and the corresponding predictions are signalled in low-frequency activity (&lt;30 Hz). Significance statement Our perception is not only determined by the information our eyes/retina and other sensory organs receive from the outside world, but strongly depends also on information already present in our brains like prior knowledge about specific situations or objects. A currently popular theory in neuroscience, predictive coding theory, suggests that this prior knowledge is used by the brain to form internal predictions about upcoming sensory information. However, neurophysiological evidence for this hypothesis is rare – mostly because this kind of evidence requires making strong a-priori assumptions about the specific predictions the brain makes and the brain areas involved. Using a novel, assumption-free approach we find that face-related prior knowledge and the derived predictions are represented and transferred in low-frequency brain activity.
BibTeX
@misc{brodskiguerniero2016information,
  author = {Alla Brodski‐Guerniero and Georg-Friedrich Paasch and Patricia Wollstadt and İpek Özdemir and Joseph T. Lizier and Michael Wibral},
  title = {{Information theoretic evidence for predictive coding in the face processing system}},
  year = {2016},
  doi = {10.1101/089300},
  howpublished = {bioRxiv},
}
Figure