Neural code uses self-information principle to organize the brain’s universal cell-assembly coding

Meng Li, Kun Xie, Hui Kuang, Jun Liu, Deheng Wang, Grace E. Fox, Zhifeng Shi, Liang Chen, Zhao Fang, Ying Mao, Joe Z. TsienbioRxiv 2017 · DOI

TL;DR

This paper proposes and validates a novel "Neural Self-Information Theory" suggesting that the brain's neural code is organized by the inherent information content of inter-spike intervals (ISIs), rather than external cues. They developed a new decoding strategy (VCAD) to identify cell assemblies based on this principle, demonstrating its effectiveness across various brain regions and behavioral tasks.

What Problem Does This Paper Address?

The brain's ability to generate complex cognition and behavior relies on how neurons communicate through their firing patterns. However, a major puzzle is how this communication (the "neural code") works in real-time, given the significant variability in neuron firing across trials, even when presented with the same stimulus. This variability has traditionally been viewed as "noise" that obscures the signal. The authors aim to understand the organizing principle behind this real-time neural code, specifically questioning if it's intrinsically generated by the neurons themselves or requires external interpretation.

Methods

The core of the paper is the Variability-based Cell-Assembly Discovery (VCAD) method. This novel approach has three main steps:

  1. Converting ISI variations into real-time variability surprisals: They model the distribution of inter-spike intervals (ISIs) using a gamma distribution. ISIs that are less probable (shorter or longer than typical) are considered "surprisals" and carry more information. These are then converted into a ternary code (positive surprisal, ground state, negative surprisal).
  2. Uncovering joint ISI surprisal patterns in space and time: Blind Source Separation (BSS) methods, specifically Independent Component Analysis (ICA), are used to identify patterns of these surprisals across simultaneously recorded neurons and over time.
  3. Identifying corresponding cell assemblies: The independent components identified by ICA represent cell assemblies. The authors use a criterion based on minimizing the coefficient of variation (CV) of neuron weights to select the most relevant ensemble patterns.

Datasets

The study utilized datasets from freely-behaving mice, recorded from multiple brain regions including:

  • Basolateral amygdala (BLA)
  • Hippocampal CA1
  • Dorsal striatum (STR)
  • Retrosplenial Cortex (RSC)
  • Prelimbic cortex (PRL)
  • Anterior cingulate cortex (ACC)

These datasets were collected during various conditions:

  • Sleep cycles
  • Fearful experiences (earthquake, elevator-drop, foot-shock)
  • Navigation tasks (linear track)
  • Operant conditioning tasks (five-choice visual discrimination)

Key Insights and Ideas

  • Neuronal variability is not just noise, but a source of information: The paper challenges the traditional view of neuronal variability as random noise. Instead, it proposes that the probability of an inter-spike interval (ISI) carries intrinsic information. Less probable ISIs (surprisals) are more informative.
  • Self-information principle as a universal code: The authors posit that this "self-information" principle is intrinsic to neurons and doesn't require an external observer to define inputs. This offers a new framework for understanding how the brain encodes information.
  • Cell assemblies are built from these intrinsic surprisals: Temporally coordinated "surprisals" across groups of neurons (cell assemblies) are proposed to form the basis of real-time neural coding.
  • Pareto Principle in neural coding: A striking finding is that approximately 20% of the ISIs (the "tail" of the distribution) are responsible for generating robust cell assembly codes, aligning with the 80/20 rule.

Key Results

  • The VCAD method successfully identified 15 distinct cell assemblies across various brain regions and behavioral tasks.
  • Neuronal variability remained consistent across different brain regions, contradicting the "system noise" model where variability would increase with information transmission.
  • Neuronal variability significantly decreased under anesthesia, supporting the idea that it's linked to cognitive processing.
  • The identified cell assemblies were time-locked to specific events, such as fearful stimuli, sleep oscillations, navigation phases, and operant conditioning stages.
  • The composition of cell assemblies (e.g., proportion of excitatory vs. inhibitory neurons, positive vs. negative surprisals) varied depending on the neural circuit, cognitive state, and behavioral task, suggesting task-specific coding strategies.

Limitations and Future Work

  • Defining "best relevant ensemble patterns": The paper introduces a criterion for selecting the optimal number of independent components (p) based on minimizing the coefficient of variation (CV) of large-contributed neurons. While this is presented as an unbiased method, the authors acknowledge that the relevance of each ensemble pattern ultimately depends on the nature of the data being analyzed.
  • Further investigation of cell subtypes: While the study identified cell assemblies, a deeper analysis of the specific roles of different neuron subtypes (e.g., O-LM cells, bursty cells, pyramidal cells) within these assemblies is suggested for future work.
  • Manipulating rewards: The study on navigation did not involve manipulating rewards, and future investigations could explore how rewards influence these cell assembly properties.
  • Understanding the "ground state": While the ground state ISIs carry less self-information, they are noted to consume energy and play a role in enabling rapid responses and ternary coding. Further exploration of their precise function is warranted.

Summary updated

Abstract

Abstract The brain generates cognition and behavior through firing changes of its neurons, yet, with enormous firing variability, the organizing principle underlying real-time neural code remains unclear. Here, we test the Neural Self-Information Theory that neural code is constructed via the self-information principle under which each inter-spike-interval (ISI) is inherently self-tagged with discrete information based on its relation to ISI variability-probability distribution - higher-probability ISIs, which reflect the balanced excitation-inhibition ground state, convey minimal information, whereas lower-probability ISIs, which signify statistical surprisals, carry more information. Moreover, temporally coordinated ISI surprisals across neural cliques intrinsically give rise to real-time cell-assembly neural code. As a result, this self-information-based neural coding is uniquely intrinsic to the neurons themselves, with no need for outside observers to set any reference point to manually mark external or internal inputs. Applying this neural self-information concept, we devised an unbiased general decoding strategy and successfully uncovered 15 distinct cell-assembly patterns from multiple cortical and hippocampal circuits associated with different sleep cycles, earthquake, elevator-drop, foot-shock experiences, navigation or various actions in five-choice visual-discrimination operant-conditioning tasks. Detailed analyses of all 15 cell assemblies revealed that ~20% of the skewed ISI distribution tails were responsible for the emergence of robust cell-assembly codes, conforming to the Pareto Principle. These findings support the notion that neural coding is organized via the self-information principle to generate real-time information across brain regions, cognitive modalities, and behaviors.
BibTeX
@misc{li2017neural,
  author = {Meng Li and Kun Xie and Hui Kuang and Jun Liu and Deheng Wang and Grace E. Fox and Zhifeng Shi and Liang Chen and Zhao Fang and Ying Mao and Joe Z. Tsien},
  title = {{Neural code uses self-information principle to organize the brain’s universal cell-assembly coding}},
  year = {2017},
  doi = {10.1101/201301},
  howpublished = {bioRxiv},
}
Figure