TL;DR
This review paper bridges the gap between microscopic neural recordings (individual neurons) and macroscopic neuroimaging (brain-wide signals) to understand how neural populations encode information. It argues that combining these approaches is crucial for a complete picture of neural processing, revealing how local network dynamics interact with global brain states to shape perception and behavior.
What Problem Does This Paper Address?
Understanding how the brain processes information has traditionally been approached from two main angles: studying the activity of individual neurons (microscopic) or observing large-scale brain activity (macroscopic, e.g., fMRI, EEG). While microscopic studies reveal details about neuronal response properties, timing, and correlations, macroscopic methods show how brain states and long-range connections influence these local activities. The challenge is to integrate these disparate levels of observation to form a cohesive understanding of how neural populations, from local circuits to distributed networks, collectively represent and process information to drive behavior. This paper aims to synthesize recent progress in this integration.
Methods
The paper is a review, so it doesn't present new experimental methods. Instead, it discusses how existing and emerging experimental and computational techniques are being used to bridge the microscopic and macroscopic divide:
- Multi-neuron recordings: High-density recordings allow simultaneous observation of hundreds of neurons, providing insights into local network dynamics, heterogeneity, sparseness, timing, and correlations.
- Neuroimaging (Mass signals): Techniques like EEG, MEG, LFP, and ECOG capture aggregate neural activity. While lacking cellular resolution, they offer excellent temporal resolution and broad spatial coverage, revealing long-range coupling and brain states.
- Computational Approaches: Advanced single-trial analysis methods are crucial for extracting meaningful information from the complex, multivariate data generated by both recording techniques. This includes decoding approaches and information theory measures (Shannon and Fisher information).
- Comparative Analysis: Studies that directly compare information extracted from mass signals and spiking activity, either through simultaneous recordings or by analyzing similar tasks across different modalities, are highlighted as key to understanding the relationship between the two.
Datasets
Not discussed in the paper. The paper reviews existing research and does not present new experimental data or datasets.
Key Insights and Ideas
- Heterogeneity is Key: Individual neurons within a population have diverse response properties (tuning curves, selectivity). This heterogeneity is not just noise; it's a fundamental ingredient that allows populations to encode a wider range of information and provides complementary information across neurons.
- Beyond Firing Rate: While firing rate is a basic code, spike timing, relative timing between neurons, and even periods of silence (absence of firing) carry crucial, often complementary, information. This is particularly relevant for understanding how information is processed at millisecond timescales.
- Mixed Selectivity: In higher brain areas, neurons often respond to combinations of multiple stimulus or task variables in complex, sometimes nonlinear ways. This "mixed selectivity" increases the dimensionality of neural representations, allowing for richer coding and easier decoding by downstream areas.
- Network State Matters: Neural responses are not solely determined by sensory input. Ongoing brain states (e.g., fluctuations in excitability, attention levels) significantly influence population activity and information processing. Mass signals are particularly useful for characterizing these global states.
- Phase is Informative: The phase of low-frequency oscillations in mass signals (like EEG/LFP) appears to be a critical carrier of sensory information, reflecting local network excitability and temporal dynamics. This phase information can be complementary to information encoded in firing rates.
- Multiplexing of Information: Different frequency bands within mass signals can carry distinct, complementary information about stimulus features, suggesting a sophisticated way the brain encodes information across multiple temporal scales.
Key Results
- Information Scaling Artifact: Early hypotheses suggested information scales linearly with population size. However, recent work indicates that this is often an artifact of random subsampling; a small, highly informative subset of neurons can carry most of the relevant information.
- Spike Timing vs. Rate: Information encoded in precise spike timing patterns is often complementary to information encoded in firing rates, and cannot be fully replaced by it.
- Phase vs. Power: In many sensory systems, the phase of low-frequency oscillations (e.g., <12 Hz) carries more sensory information than the power of those oscillations.
- Correlation Impact: The precise pattern of noise correlations (trial-to-trial variability) matters more than its strength. Certain types of correlations ("differential correlations") can limit information encoding, even with larger populations.
Limitations and Future Work
- Bridging the Gap: A major challenge remains in fully integrating insights from microscopic and macroscopic measurements. How do specific cell types contribute to mass signals? How do global brain states (seen in mass signals) relate to local coding properties?
- Experimental Challenges: While multi-neuron recordings are advancing, capturing activity from all relevant neurons in a population remains difficult. Similarly, interpreting mass signals requires careful consideration of the multiple neural phenomena they represent.
- Future Directions:
- Develop better statistical methods for single-trial analysis that can be applied across different measurement modalities.
- Continue comparative studies between mass signals and spiking activity to understand how information is represented at different scales.
- Investigate the role of mixed selectivity in temporal and frequency domains, not just across neurons.
- Develop more sophisticated computational models that can link neural signals to computing architectures and explain both spiking and mass activity.
- Integrate advances in cellular imaging and histochemistry to understand the contribution of specific cell types to population codes.
Summary updated