Random neural networks match observed dimensionality of neural population recordings and motivate stronger experimental tests

Problem Statement
Randomly connected neural networks have long served as a theoretical framework for studying collective dynamics in neural populations. However, quantitative comparisons between these minimal models and modern large-scale experimental recordings have remained limited. A central observation in experimental neuroscience is that neural population activity is often low-dimensional, meaning the firing patterns of many neurons vary in only a few coordinated ways rather than independently. While this low dimensionality has often been interpreted as evidence of specially organized, fine-tuned circuitry, it remains an open question whether such low dimensionality can arise generically from random recurrent interactions without fine-tuning. Previous theoretical work using Dynamical Mean-Field Theory (DMFT) calculated the dimensionality of random networks in the limit of infinite measurement time and without external inputs. This work addresses the gap between theory and experiment by determining if a minimally structured random network can quantitatively account for the low dimensionality observed in real data when realistic experimental constraints—specifically finite measurement time and external behavioral inputs—are incorporated.

Methodology
The authors employ a standard random recurrent neural network model (Sompolinsky et al.) with $N$ neurons, where the coupling matrix $J$ is sampled from a zero-mean Gaussian distribution ($J_{ij} \sim \mathcal{N}(0, g^2/N)$). To model behavioral contexts, the network receives time-independent external inputs $f_i$ sampled from a zero-mean Gaussian distribution with strength $I$. The dynamics are governed by a continuous-time rate model with a nonlinearity $\phi(h) = \text{erf}(\sqrt{\pi}h/2)$.

The core theoretical tool is Dynamical Mean-Field Theory (DMFT), extended to handle:

1. External Inputs: Decomposing activity into an "ordered response" ($\bar{r}$), which is the time-averaged response to the input, and "temporal chaos" ($\tilde{r}$), representing the residual fluctuations.
2. Finite Measurement Time ($T$): Deriving analytical expressions for statistics measured over a finite window $T$, acknowledging that the measured dimensionality depends on the ratio of measurement time to the autocorrelation time of the chaos ($\tau_{\tilde{C}}$).
3. Multi-Context Analysis: Using replica methods to calculate statistics across multiple behavioral contexts (different external inputs) to assess the geometry of the "cloud" of ordered responses.

Key quantities calculated include:

• Participation Ratio (PR): Used as a linear measure of dimensionality for both temporal chaos and ordered responses.
• Cosine Similarity (CS): Measuring the similarity between ordered responses across different contexts.
• Orientation Similarity (OS): Measuring the similarity between the principal axes (covariance ellipsoids) of temporal chaos across contexts.

Key Contributions and Results

• Quantitative Agreement with Experiments: The authors show that when finite measurement time and external inputs are included, the dimensionality predicted by a minimally structured random network is quantitatively consistent with experimental data from primate motor cortex (PMd and M1) during reaching tasks. Specifically, the model reproduces the low dimensionality values observed in experiments for realistic measurement durations ($T \sim 1$ s) and neuron counts ($N \sim 800$).
• Limitations of Dimensionality as a Discriminator: Despite the agreement, the authors find that in the regime defined by current experimental durations, the predicted dimensionality is low and relatively insensitive to network parameters (such as gain $g$ or input strength $I$). Consequently, dimensionality measurements alone cannot distinguish between a minimally structured random network and a network with additional specific wiring structures. The agreement may simply reflect the limited measurement window rather than the absence of structure.
• Non-Monotonic Dimensionality of Temporal Chaos: The long-time dimensionality of temporal chaos ($PR_\infty$) varies non-monotonically with external input strength. It initially increases as weak inputs align fluctuations with high-variance directions of the autonomous chaos, then decreases as strong inputs drive neurons into low-gain regions of the nonlinearity, suppressing chaos.
• Differential Sensitivity of Geometric Measures:
  • Ordered Response Similarity (CS): The similarity between time-averaged responses across contexts remains high and saturates quickly as input strength increases, preserving the similarity of the inputs. This metric is primarily sensitive to single-neuron properties.
  • Temporal Chaos Orientation Similarity (OS): In contrast, the orientation of the fluctuation ellipsoid (temporal chaos) changes rapidly with input strength. Even when ordered responses are highly similar, the underlying fluctuation structures can decorrelate significantly. This metric is more sensitive to network structure than dimensionality.
  • Multi-Context Dimensionality: The dimensionality of the cloud of ordered responses across multiple contexts grows systematically with the number of contexts sampled. This provides a baseline against which experimental data can be compared to detect deviations from the minimal model.

Significance
The paper argues that low dimensionality in neural population recordings should not be taken as definitive evidence for specially organized circuit structure. A minimally structured random network, when subjected to realistic experimental constraints (finite time and external inputs), naturally produces activity with low dimensionality comparable to experimental observations. Therefore, the mere observation of low dimensionality is insufficient to infer the presence of fine-tuned connectivity.

To resolve whether a network possesses structure beyond random interactions, the authors propose that experiments must move beyond simple dimensionality counts. They identify specific geometric quantities that provide stronger tests:

1. The non-monotonic dependence of temporal chaos dimensionality on input strength.
2. The rapid decay of orientation similarity (OS) between temporal chaos structures across different behavioral contexts.
3. The scaling of multi-context dimensionality with the number of sampled contexts.

Deviations from the quantitative predictions of the minimally structured random network for these specific measures would indicate the presence of additional, non-random organization in the underlying neural circuitry. The work provides a quantitative framework for designing experiments to distinguish between generic random dynamics and structured computations.