Strong information-limiting correlations in early visual areas

Jorrit S. Montijn, Rex G. Liu, Amir Aschner, Adam Kohn, Peter E. Latham, Alexandre PougetbioRxiv 2019 · DOI

TL;DR

This paper introduces a novel method to quantify "differential correlations" in neural activity, a key factor limiting information transfer in the brain. They demonstrate that in macaque V1, neural information aligns closely with behavioral performance, suggesting efficient downstream processing.

What Problem Does This Paper Address?

A fundamental question in neuroscience is how efficiently neural circuits process information. If processing is efficient, the information available in early sensory areas should closely match an animal's behavioral performance. However, despite decades of research, it's been unclear how efficient the brain truly is. A major hurdle has been quantifying "differential correlations" – a specific type of noise in neural activity that can significantly limit the amount of information encoded, especially in large populations of neurons. Detecting and measuring these correlations has been technically challenging.

Methods

The authors developed a new method to detect and quantify differential correlations. This method relies on analyzing the eigenspectrum of the noise covariance matrix of neural activity. Specifically, they look at how well the neural stimulus encoding direction (f') aligns with the eigenvectors of this covariance matrix.

Here's a breakdown:

  • The Core Idea: In the presence of strong differential correlations, the encoding direction (f') should be strongly aligned with the eigenvectors corresponding to the largest eigenvalues of the noise covariance matrix.
  • The Metric (φ): They define a metric, denoted by φ, which quantifies this alignment. It's essentially the difference between the alignment of f' with the principal components of the real noise covariance matrix and the alignment with the principal components of a shuffled covariance matrix (where temporal correlations are destroyed). A higher φ value indicates stronger differential correlations.
  • Computational Models: To validate their method, they used two biologically plausible computational models of the early visual system (LGN-V1 circuit). These models allowed them to simulate neural activity with controlled levels of differential correlations.
  • Experimental Data Analysis: They applied their method to electrophysiological recordings from the primary visual cortex (V1) of awake macaque monkeys.

Datasets

The study utilized two main sources of data:

  • Simulated Data: Data generated from two computational models of the LGN-V1 circuit. These models allowed for controlled manipulation of differential correlations.
  • Experimental Data: Electrophysiological recordings from the primary visual cortex (V1) of three awake macaque monkeys, obtained using Utah arrays. Seven data sets were analyzed.

Key Insights and Ideas

  • Differential Correlations are Key: The paper strongly emphasizes that differential correlations, not just random noise, are the primary bottleneck for information transfer in large neural populations.
  • A New Way to Measure Information Bottlenecks: The proposed method provides a statistically robust way to infer the strength of these information-limiting correlations, even with a moderate number of neurons and trials.
  • V1 is Surprisingly Efficient: The findings suggest that V1 in macaque monkeys holds a significant amount of information, and this information is closely matched by behavioral performance. This implies that downstream circuits are quite good at extracting this information.
  • Behavioral Relevance: The method translates the measured differential correlations into an equivalent orientation discrimination threshold, directly linking neural processing to observable behavior.

Key Results

  • Method Validation: Simulations showed that their metric (φ) accurately detected differential correlations and that these correlations scaled with the strength of injected noise, consistent with theoretical predictions.
  • Model Sensitivity: The method proved robust across different model architectures and noise injection strategies.
  • Experimental Findings: In four out of seven V1 data sets from awake monkeys, they found statistically significant evidence of differential correlations.
  • Neural vs. Behavioral Information: The estimated neural discrimination thresholds derived from their analysis (0.47-1.20 degrees) were comparable to typical behavioral thresholds (1-2 degrees). This suggests that V1 contains nearly as much information as is utilized in behavior.
  • Information Saturation: The results argue against extreme information saturation at very high levels, as suggested by some previous theories, and instead support a scenario where neural information is close to behavioral information.

Limitations and Future Work

  • Recording Quality: The authors acknowledge that in some experimental data sets, the value of φ was not significantly different from zero. They hypothesize this might be due to the degradation of recording quality in chronic arrays over time, rather than a true absence of differential correlations.
  • Suboptimal Computation: While V1 appears efficient, information loss still occurs between V1 and behavior. The authors suggest this could be due to suboptimal computation in downstream areas, such as inefficient integration of information across space and time, or suboptimal synaptic weights between cortical areas.
  • Future Directions: The paper opens avenues for further research into the precise mechanisms of information loss downstream of V1 and how these differential correlations might be modulated by factors like attention or learning. The method itself could be applied to other sensory areas or cognitive tasks to assess information processing efficiency across the brain.

Summary updated

Abstract

Abstract If the brain processes incoming data efficiently, information should degrade little between early and later neural processing stages, and so information in early stages should match behavioral performance. For instance, if there is enough information in a visual cortical area to determine the orientation of a grating to within 1 degree, and the code is simple enough to be read out by downstream circuits, then animals should be able to achieve that performance behaviourally. Despite over 30 years of research, it is still not known how efficient the brain is. For tasks involving a large number of neurons, the amount of information encoded by neural circuits is limited by differential correlations. Therefore, determining how much information is encoded requires quantifying the strength of differential correlations. Detecting them, however, is difficult. We report here a new method, which requires on the order of 100s of neurons and trials. This method relies on computing the alignment of the neural stimulus encoding direction, f′ , with the eigenvectors of the noise covariance matrix, Σ . In the presence of strong differential correlations, f′ must be spanned by a small number of the eigenvectors with largest eigenvalues. Using simulations with a leaky-integrate-and-fire neuron model of the LGN-V1 circuit, we confirmed that this method can indeed detect differential correlations consistent with those that would limit orientation discrimination thresholds to 0.5-3 degrees. We applied this technique to V1 recordings in awake monkeys and found signatures of differential correlations, consistent with a discrimination threshold of 0.47-1.20 degrees, which is not far from typical discrimination thresholds (1-2 deg). These results suggest that, at least in macaque monkeys, V1 contains about as much information as is seen in behaviour, implying that downstream circuits are efficient at extracting the information available in V1.
BibTeX
@misc{montijn2019strong,
  author = {Jorrit S. Montijn and Rex G. Liu and Amir Aschner and Adam Kohn and Peter E. Latham and Alexandre Pouget},
  title = {{Strong information-limiting correlations in early visual areas}},
  year = {2019},
  doi = {10.1101/842724},
  howpublished = {bioRxiv},
}
Figure