TL;DR
This paper investigates how the brain communicates information within the cortex, specifically testing two competing "predictive coding" theories. Using information theory, they found that the brain generally favors transmitting predictable information (reliability coding) over unpredictable information (error coding), though both strategies appear to be employed.
What Problem Does This Paper Address?
The brain is thought to constantly predict incoming sensory information using internal models. When predictions don't match reality, this "prediction error" needs to be communicated. However, there's a debate: do neurons send the unpredictable part of the signal (error coding) or the predictable part (reliability coding) up the cortical hierarchy? This paper aims to resolve this by examining how neural predictability and information transfer are correlated in cortical connections.
Methods
The authors employed an information-theoretic approach using two key measures:
- Active Information Storage (AIS): Quantifies how much past neural activity predicts current activity, serving as a proxy for neural predictability.
- Transfer Entropy (TE): Measures the directed flow of information from one neural source to a target.
They then calculated the Local Storage-Transfer Correlation (LSTC), which is the correlation between AIS and TE for a given connection. The sign of this correlation is crucial:
- Positive LSTC: Indicates that as predictability (AIS) increases, information transfer (TE) also increases, suggesting reliability coding (transmitting predictable information).
- Negative LSTC: Indicates that as predictability (AIS) increases, information transfer (TE) decreases, suggesting error coding (transmitting unpredictable information).
They used both bivariate and multivariate TE analyses to infer connectivity.
Datasets
The study utilized two datasets:
- Human intracranial EEG (iEEG) data: Recorded from 9 epileptic patients performing a face-processing task. This dataset allowed for analysis of LFP (local field potential) activity.
- Mouse visual cortex spike data: Obtained from the Allen Brain Observatory, recording neuronal spiking activity during a drifting grating visual stimulus.
Key Insights and Ideas
- Predictive coding isn't a single strategy: The brain appears to employ both reliability and error coding mechanisms within its cortical connections.
- Reliability coding dominates: Across both human and mouse data, there's a strong tendency for information transfer to increase with neural predictability. This suggests the brain often prioritizes sending along what it expects.
- Resolution matters: While human LFP data (coarse-grained) exclusively showed positive LSTCs (reliability coding), mouse spike data (fine-grained) revealed both positive and negative LSTCs, indicating that error coding might be more apparent at the single-neuron level.
- "Explaining away" at the neural level: The negative LSTCs observed in mouse data align with the concept of "explaining away," where predictable inputs lead to reduced information transfer, possibly by suppressing expected signals.
Key Results
- Human iEEG: All 17 detected cortico-cortical connections with significant transfer entropy showed positive LSTCs, indicating a predominant reliability coding strategy across frontal, parietal, and temporal regions.
- Mouse visual cortex: Out of 9 significant connections, 66.67% showed positive LSTCs (reliability coding), while 33.33% showed negative LSTCs (error coding).
Limitations and Future Work
- Resolution sensitivity: The authors acknowledge that the LSTC measure's sensitivity might depend on the resolution of neural recordings (LFP vs. spikes). Future work should systematically explore this.
- Stimulus predictability vs. neural predictability: While the study focused on neural predictability, future research could further investigate how external stimulus predictability interacts with these internal coding strategies.
- Limited mouse data: The mouse spike data analysis is considered an "exploratory study" due to challenges in detecting single-unit interactions and the complexity of neuronal inputs.
- "Explaining away" nuances: The authors suggest that the absence of negative LSTCs in LFP data might be due to averaging over many neurons, masking the finer-grained error coding signals. Further investigation into how these signals propagate and interact across different levels of the hierarchy is warranted.
Summary updated