Low-dimensional models of neural population activity in sensory cortical circuits

Evan W. Archer, Urs Köster, Jonathan W. Pillow, Jakob H. Macke

Understanding how the brain processes sensory information requires deciphering the activity of large, interconnected populations of neurons. Researchers have recently developed the Quadratic Input Latent Dynamical System (QLDS), a statistical framework that simultaneously captures how sensory stimuli drive neural populations and how internal brain dynamics shape those responses. By combining nonlinear "receptive fields" with a low-dimensional dynamical model, QLDS provides a compact, scalable way to understand the collective behavior of hundreds of neurons in the visual cortex.

The "Great Divide" in Neural Modeling

In computational neuroscience, there has traditionally been a divide in how neural activity is modeled:

  1. System Identification: This approach focuses on how a single neuron responds to a stimulus (e.g., a specific image pattern), treating each cell as an independent unit.
  2. Latent Variable Models: This approach focuses on the internal "chatter" or dynamics of a population of neurons, often ignoring the external stimulus or treating its influence as purely linear.

The reality is that neurons in the cortex are highly interconnected; they share inputs and influence each other’s dynamics. Analyzing them one by one misses the "population code," while ignoring the stimulus makes it impossible to understand how the brain actually processes information. Because neural data is high-dimensional and noisy, modeling both stimulus-driven and internally-driven activity simultaneously has historically been computationally prohibitive.

The QLDS Framework

To bridge this gap, the authors introduce the QLDS model, which integrates three main components into a single cohesive architecture:

The QLDS architecture
  1. Nonlinear Stimulus Input: Instead of assuming neurons simply sum up visual inputs linearly, the model uses a quadratic function. This allows it to capture "complex cell" behaviors, such as responding to a specific orientation regardless of its exact spatial position (phase invariance).
  2. Latent Dynamics: The stimulus doesn't drive the neurons directly. Instead, it drives a small number of "latent states" (a low-dimensional hidden vector). These states evolve over time according to a linear dynamical system, capturing the "momentum" and recurrent connections of the neural circuit.
  3. Poisson Observation Model: To account for the discrete, "spiky" nature of neural data, the model assumes the observed spikes are generated by a Poisson process influenced by the latent states.

To make this mathematically intensive model practical, the authors developed a fast estimation method using Expectation-Maximization (EM) combined with Laplace approximations. This ensures that the time it takes to fit the model grows linearly—not quadratically—with the number of neurons and the length of the recording.

Key Insights and Advantages

The QLDS model offers several distinct advantages over traditional approaches:

  • The "Bottleneck" Advantage: Because the model assumes a low-dimensional latent space, it is highly parsimonious. Instead of needing a unique filter for every single neuron, it identifies a few shared filters that drive the entire population.
  • Capturing Shared Variability: The model naturally explains why neurons often fire together even when there is no stimulus (noise correlations). These correlations emerge because the neurons are all "listening" to the same underlying latent dynamical system.
  • Scalability: The linear scaling of the inference algorithm is a major win for modern neuroscience, where researchers are increasingly recording from thousands of neurons simultaneously.

Experimental Validation and Results

The model was rigorously validated on three distinct types of data: simulated data (to prove the model could accurately recover hidden "ground truth" filters), multi-channel recordings from Cat V1 (primary visual cortex), and recordings from 113 neurons in an anesthetized Macaque V1 responding to drifting grating stimuli.

Model fit vs. true correlations
  • Subspace Recovery: In simulations, QLDS successfully identified the specific visual features that drove the population, even when the latent states were mathematically rotated or transformed.
  • Correlation Matching: In the cat V1 data, the model's synthetic data closely matched the actual cross-neural correlations found in the real biological data. It successfully captured "burst-like" events where many neurons fired synchronously.
  • Tuning Curves: For the macaque data, QLDS accurately predicted the orientation tuning of the neural population, modeling both the average response to a moving bar and the trial-to-trial variability.
  • Efficiency: The model achieved high accuracy with far fewer parameters than "fully-connected" models (like Generalized Linear Models), which typically require parameters to scale with the square of the number of neurons.

Limitations and Future Directions

While powerful, the QLDS framework has limitations. Like many latent variable models, the internal states suffer from identifiability issues; they can be mathematically "rotated" without changing the output, meaning specific latent dimensions don't always have a direct, one-to-one biological interpretation. Additionally, while quadratic models are excellent for V1, higher-level brain areas (like V4 or IT) likely require deeper nonlinearities to describe their stimulus preferences.

Finally, the reliance on a Gaussian (Laplace) approximation for the posterior is fast but might be less accurate for neurons with very low firing rates. Future work could extend this framework to include more complex stimulus models, such as deep neural networks, while retaining the highly efficient latent dynamical core.

Summary updated

Abstract

Neural responses in visual cortex are influenced by visual stimuli and by ongoing spiking activity in local circuits. An important challenge in computational neuroscience is to develop models that can account for both of these features in large multi-neuron recordings and to reveal how stimulus representations interact with and depend on cortical dynamics. Here we introduce a statistical model of neural population activity that integrates a nonlinear receptive field model with a latent dynamical model of ongoing cortical activity. This model captures the temporal dynamics, effective network connectivity in large population recordings, and correlations due to shared stimulus drive as well as common noise. Moreover, because the nonlinear stimulus inputs are mixed by the ongoing dynamics, the model can account for a relatively large number of idiosyncratic receptive field shapes with a small number of nonlinear inputs to a low-dimensional latent dynamical model. We introduce a fast estimation method using online expectation maximization with Laplace approximations. Inference scales linearly in both population size and recording duration. We apply this model to multi-channel recordings from primary visual cortex and show that it accounts for a large number of individual neural receptive fields using a small number of nonlinear inputs and a low-dimensional dynamical model.
BibTeX
@inproceedings{archer2014low,
  author = {Evan W. Archer and Urs Köster and Jonathan W. Pillow and Jakob H. Macke},
  title = {{Low-dimensional models of neural population activity in sensory cortical circuits}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2014},
}
Figure