TL;DR
This paper demonstrates that the trial-to-trial variability in mouse visual cortex (V1) neurons is primarily explained by the activity of their local "functional groups" – sets of interconnected neurons exhibiting correlated activity – rather than global factors like stimulus or locomotion. This finding suggests that understanding these local recurrent interactions is crucial for a comprehensive theory of sensory encoding.
What Problem Does This Paper Address?
Understanding how the brain processes sensory information is a fundamental challenge. A key hurdle is the inherent variability in neuronal responses, even when presented with identical stimuli. While global factors like visual input and animal behavior (e.g., running speed) are known to influence neuronal activity, it's unclear how much of this trial-to-trial variability they truly account for. Previous research has hinted that local interactions between neurons might play a more significant role than previously thought. This paper aims to disentangle the contributions of local neuronal interactions versus global factors in explaining single-trial response variability in V1.
Methods
The researchers employed a linear modeling (LM) approach to predict the fluorescence activity of individual V1 neurons. They defined "functional groups" for each neuron based on pairwise partial correlations of their activity, after accounting for stimulus and running speed. These functional weights, representing the strength and direction of these correlations, were then used as input features in the LM. To assess the importance of different factors, they compared the predictive performance of a full LM (including functional group activity, stimulus, and running speed) against restricted LMs where one or more of these factors were excluded. They also explored the temporal dynamics of these interactions by analyzing correlations at different time lags and investigated the role of "strong" weights and edges (those in the top quartile of magnitudes) through permutation analyses. Finally, they used a Bayesian decoding framework to assess how well this functional group information could be used to decode visual stimuli.
Datasets
The study utilized calcium imaging data from L2/3 excitatory neurons in the visual cortex of awake, freely running mice. This data was previously published in Dechery & MacLean (2018) and consists of recordings from 8 animals across 21 distinct fields of view. The mice viewed drifting square-wave gratings presented in 12 directions, interleaved with grey screen periods.
Key Insights and Ideas
- Local is (mostly) king: The activity of a neuron's immediate neighbors (its "functional group") is a far better predictor of its own trial-to-trial activity than global factors like the visual stimulus or the animal's running speed.
- The Hebbian assembly connection: The concept of "functional groups" provides an operational definition for Hebbian assemblies, suggesting that these coactive groups are fundamental units of sensory representation.
- Short timescales matter most: The most informative correlations within functional groups occur over very short timescales (around 30ms), highlighting the importance of rapid, local interactions.
- Positive correlations drive prediction: While both positive and negative correlations exist, it's the strong positive correlations within functional groups that are crucial for accurate predictions of neuronal activity.
- Topology and weight precision are key: The specific connections (edges) and their precise strengths (weights) within the functional group are critical for accurate predictions, not just the general presence of strong connections.
Key Results
- Including functional group activity in an LM improved single-trial prediction accuracy significantly more than models relying solely on stimulus or running speed.
- Restricting the LM to exclude functional group coupling led to a substantial increase in prediction error (median increase of ~29% in test set MSE).
- The most informative correlations were strong, positive, and occurred at lag 0 (bidirectional, ~30ms).
- Excluding lag 0 coupling (bidirectional interactions) drastically reduced prediction accuracy, while excluding lag 1 coupling had a minimal effect.
- Decoding visual stimuli using the functional group information improved accuracy by approximately 64% compared to a decoder assuming neuronal independence.
- Permuting strong weights or edges within functional groups significantly degraded prediction accuracy, underscoring the importance of specific local connectivity.
Limitations and Future Work
The authors acknowledge that their linear model is a simplification and that more complex, nonlinear models might capture additional dynamics. They also note that while their findings strongly suggest the computational relevance of functional groups for downstream circuits, their study doesn't definitively prove that these circuits actually access or utilize this information. Future work could explore how these functional groups are formed and modified, investigate their role in more complex computations beyond simple stimulus decoding, and examine these dynamics in other brain areas. The study also highlights the need for further research into the precise timescales and computational capabilities of these neuronal assemblies.
Summary updated