Imagine trying to understand the dynamics of a crowded marketplace by listening only to the three vendors with megaphones. You would likely conclude that the entire market is loud, highly coordinated, and focused exclusively on selling fish.
This is the fundamental challenge of modern neuroscience and artificial intelligence interpretability. Whether recording from billions of biological neurons or analyzing billions of artificial parameters, researchers rarely observe the full system. Instead, they work with a tiny, sparse subset of units.
Crucially, this sparse sampling is not random; it is systematically distorted. The tools we use to observe neural networks act as filters, creating profound selection biases that warp our understanding of population dynamics, network geometry, and computational architecture.
The Dimensions of Selection Bias
Experimental and analytical tools introduce systematic filters across multiple dimensions: activity levels, physical morphology, and genetic expressions.
Activity Bias: The Silent Majority
In electrophysiology and optical imaging, the most glaring distortion is the over-representation of highly active units. As Brecht et al. point out, standard recording methods favor neurons with high firing rates, effectively ignoring a "silent majority" of cells that fire infrequently. This activity bias means that current models of neural coding are built on an unrepresentative, highly active subset of the population (Carandini et al.). When characterizing population dynamics or representational stability over time, favoring these highly active units can lead to conflicting or skewed conclusions about phenomena like neural drift (Bonhoeffer et al.; Dąbrowska et al.).
A striking parallel exists in artificial neural networks. In mechanistic interpretability, researchers often isolate the function of a neuron by looking at the inputs that cause the highest activation. However, Garde et al. argue that prioritizing maximum activation events fails to capture the subtle, crucial functional roles that neurons perform at lower or average activation thresholds. By focusing only on the "top-activating" inputs, we ignore inhibitory signals and lower-threshold behaviors, resulting in an incomplete profile of the model's true operational state (Bai et al.).
Physical and Genetic Gatekeeping
Biases also emerge from the physical and genetic properties of the cells themselves. In patch-clamp recordings, electrodes are physically guided toward cell bodies. This creates a size-dependent sampling bias: smaller neurons are harder to target and stabilize, leading to their severe underrepresentation in functional datasets (Tavakoli et al.).
Similarly, labeling methods that rely on specific transgenic lines or retrograde viral tracers to isolate neurons based on their downstream targets introduce a strict selection bias (Zhang et al.; Svoboda et al.). By filtering for a single projection pathway, researchers risk missing the broader functional diversity of the region, overlooking subpopulations that channel entirely different computational logic to other targets (Feng et al.; Reddy et al.).
How Subsampling Distorts Network Geometry
When we observe only a fraction of a high-dimensional system, the geometric and statistical properties of the network do not just lose resolution—they fundamentally change.
Spurious Correlations vs. Latent Drivers
If you record from two connected neurons, it is tempting to assume their correlated activity implies direct functional coupling. However, Macke et al. demonstrate that sparse sampling makes it mathematically difficult to distinguish between direct interactions and the shared influence of unobserved neighbors.
True Network (Unobserved Latent Hub) Observed Subsample (Spurious Coupling)
[Hidden Neuron]
/ \
v v
[Neuron A] <---> [Neuron B] [Neuron A] <=======> [Neuron B]
Because we only record a small fraction of the network, the true high-dimensional structure is obscured (Representational Dissimilarity...). What looks like an intrinsic "tuning curve"—a neuron dedicated to a specific external stimulus—is often an artifact of internal network dynamics and unobserved common drivers that have been "explained away" by the missing population (Stevenson et al.).
Blinding Architectural Metrics
This loss of dimensionality has severe consequences when comparing biological networks to artificial ones. A common paradigm in computational neuroscience is to evaluate an AI model by checking how well its representations align with biological neural recordings.
However, Han demonstrates that sparse sampling drastically degrades the statistical power of similarity metrics. When a dataset contains only a small fraction of a system's total units, fundamentally different architectures (such as Convolutional Neural Networks versus Transformers) can yield deceptively similar representational geometries. Subsampling effectively masks the structural differences in how information is processed across the whole population.
Averaging Out Sparse Codes
When researchers sample a population randomly, they often conclude that sensory information is broadly distributed across a wide, redundant pool of cells. Ince et al. argue that this is an artifact of the sampling math. Randomly sampling and averaging population activity washes out the contributions of highly specialized, temporally precise ensembles. Instead of a broad, distributed code, the brain frequently relies on sparse, high-efficiency coding schemes that are easily missed unless the sampling strategy is explicitly designed to find them.
The Physics of Firing Rates as Sampling Frequencies
The distortion caused by sparse sampling is not just spatial or structural; it is also temporal. To understand why, we can look at the motor system, where the central nervous system delivers a "common drive" to a pool of motor neurons to control muscle force (Negro et al.).
Negro provides a elegant formulation of this problem by treating individual neurons as biological samplers. A neuron can only transmit information when it fires a spike. Therefore, its mechanical firing rate acts as a sampling frequency for its underlying synaptic inputs.
$$\text{Sampling Frequency} \approx \text{Neuronal Firing Rate}$$
If a motor neuron fires at 10 Hz, the Nyquist theorem dictates it cannot mathematically capture or transmit any components of the common synaptic drive higher than 5 Hz. Relying on individual pairs of sparsely sampled neurons creates a severe information bottleneck. The low discharge rate aliases the high-frequency components of the underlying command, providing an incomplete and biased picture of the total neural drive.
Mitigating the Sampling Bottleneck
To overcome these constraints, researchers are moving away from treating individual recorded units as isolated truth-tellers, turning instead to population-level reconstruction frameworks.
- Latent Dynamical Modeling: Rather than trying to map direct cell-to-cell coupling from sparse samples, frameworks proposed by Macke et al. treat shared variance as a hidden, low-dimensional latent signal. This shifts the goal from capturing every unit to capturing the underlying dynamical state space.
- Cumulative Spike Trains (CSTs): To solve the temporal aliasing of low firing rates, Negro demonstrates that pooling spikes from multiple neurons into a single Cumulative Spike Train effectively increases the sampling frequency. This allows researchers to accurately reconstruct the full frequency spectrum of the common synaptic input.
- Biophysical Population Reconstruction: To counteract the physical bias toward large, high-threshold units, Caillet et al. developed frameworks that combine partial experimental samples (like high-density EMG) with mathematically realistic biophysical models. By using the sparse sample to anchor a simulation of the entire neuron pool, they reconstruct the hidden activity of low-threshold units that are critical for behavior but typically invisible to recording hardware.
Ultimately, recognizing that our data is a highly filtered, non-random subset of a larger system is the first step toward accurate interpretation. By shifting our focus from individual units to latent spaces and population-wide reconstructions, we can begin to see the true structure of the network, rather than just the parts that happen to shout the loudest.
Key Foundational Work
Chronic multiunit recordings in behaving animals: advantages and limitations (2004) — contributed 2 concept elements to this direction.
Novel Approaches to Monitor and Manipulate Single NeuronsIn Vivo (2004) — contributed 1 concept elements to this direction.
Flight Control in the Dragonfly: A Neurobiological Simulation (1990) — contributed 1 concept elements to this direction.