Modeling a population of retinal ganglion cells with restricted Boltzmann machines

Riccardo Volpi, Matteo Zanotto, Alessandro Maccione, Stefano Di Marco, Luca Berdondini, Diego Sona, Vittorio MurinoScientific 2020 · DOI

TL;DR

This paper demonstrates that Restricted Boltzmann Machines (RBMs) can effectively model the complex spiking activity of retinal ganglion cells (RGCs) in an unsupervised manner. By analyzing this activity, RBMs can uncover latent patterns that directly correspond to visual stimuli, even under altered retinal conditions, and can even capture temporal dynamics.

What Problem Does This Paper Address?

The retina is our primary visual processing hub, encoding raw visual information before it's sent to the brain. Understanding how retinal ganglion cells (RGCs), the output neurons of the retina, encode this information is crucial for advancing our understanding of vision and developing technologies like retinal prosthetics. While previous work has used latent variable models to study RGC activity, this paper specifically validates the use of Restricted Boltzmann Machines (RBMs) for this task. The challenge lies in modeling the high-dimensional, complex, and often noisy joint spiking activity of a large population of RGCs in an unsupervised way, meaning without prior knowledge of the stimuli presented. The authors aim to show that RBMs can discover meaningful patterns (modes) in this activity that are directly linked to the visual input, even when the retina's processing is pharmacologically altered.

Methods

The core of the methodology revolves around using Restricted Boltzmann Machines (RBMs) to model the joint distribution of RGC firing rates. Specifically, they employ two variants:

  • Mean-Covariance Restricted Boltzmann Machines (mcRBMs): These are used to model the statistical regularities (modes) in the RGC population activity that correspond to static visual stimuli. mcRBMs are adept at capturing the mean and covariance of the observed data, allowing them to represent different modes of activity.
  • Conditional Restricted Boltzmann Machines (cRBMs): These are employed to capture the temporal dynamics of RGC responses. cRBMs extend RBMs by incorporating information from past inputs, enabling the modeling of time-varying patterns.

To analyze what the RBMs have learned, they use a spike-triggered average (STA)-like visualization technique. This involves selecting a specific latent state or unit in the RBM, retrieving the RGC firing rate samples that activate it, and then averaging the corresponding visual stimuli. This helps to visualize the visual features encoded by the RBM.

For quantitative evaluation, they compute the mutual information between the RBM's latent states and the presented stimuli. This metric quantifies how well the model has captured the relationship between the visual input and the neural activity.

Datasets

The paper utilizes datasets of RGC activity recorded from explanted mouse retinas. These recordings were performed using high-density 64x64 CMOS-based Multi-Electrode Arrays (MEAs). The stimuli presented to the retinas included:

  • Simple stimuli: Square gratings with varying orientations and spatial frequencies.
  • Complex stimuli: Natural scenes, specifically images of a brick wall with a moving window.

The experiments also involved recordings under different conditions:

  • Normal conditions: Healthy retinas.
  • Altered encoding conditions: Retinas treated with pharmacological agents (TPMPA and Bicuculline) to block GABA receptors, mimicking impaired visual information processing.

No new benchmarks were introduced; rather, existing experimental setups and data recording techniques were used.

Key Insights and Ideas

  • Unsupervised Discovery of Stimulus-Encoded Modes: The most significant insight is that RBMs can, without any prior knowledge of the stimuli, discover distinct "modes" or patterns in the RGC population activity that are directly correlated with specific visual features or stimuli. This suggests that the retina's output naturally organizes itself into statistically discernible states related to what it's seeing.
  • Physiological Relevance Under Impairment: The RBM model's ability to reflect known physiological behaviors is a key finding. When the retina's processing is impaired (e.g., by blocking GABA receptors), the model's performance (measured by mutual information) degrades, and the learned modes become less specific, sometimes even encoding stimuli in a "pathological" way (like counterphase stimuli being grouped together). This validates the model's sensitivity to the functional state of the retina.
  • Capturing Temporal Dynamics: The use of cRBMs demonstrates that these models can go beyond static patterns and capture the temporal evolution of stimuli. This is crucial for understanding how the retina processes dynamic visual information, such as motion.
  • Latent States vs. Hidden Units: The paper distinguishes between latent states (combinations of hidden unit activations) and individual hidden units. Latent states appear to encode more specific stimulus features (like gratings in a particular position), while individual hidden units seem to represent more generalized visual features, akin to receptive fields.
  • Impact of Neuron Count: The results suggest that having more recorded neurons (as seen in patch t7 vs. other patches) leads to more informative and less noisy representations, highlighting the importance of large-scale recordings for robust modeling.

Key Results

  • Experiment 1 (Heterogeneous Grating Stimulation): mcRBMs successfully identified modes in RGC activity that corresponded to different orientations of square gratings. Visualizations showed that latent states were associated with specific grating orientations.
  • Experiment 2 (GABA Blockage): Pharmacological impairment of GABA receptors led to a significant decrease in the mutual information between RBM states and stimuli, particularly for higher spatial frequencies. This confirmed that the model's ability to decode stimulus information is sensitive to the retina's functional integrity. For instance, in impaired retinas, stimuli in counterphase were encoded in the same latent state, a deviation from normal processing.
  • Experiment 3 (Natural Scenes): Even with more complex stimuli like natural scenes (brick wall images), mcRBMs could still model the population activity and retrieve distinct visual features encoded by latent variables.
  • Experiment 4 (Population Dynamics): cRBMs were able to model temporal patterns associated with stimulus dynamics, such as grating motion and black-to-white transitions. Visualizations showed how different hidden states and units captured these temporal evolutions.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that STA-based representations are slightly blurred due to averaging, though this is also seen as a positive aspect reflecting similarity in responses.
  • The performance of the model is shown to be dependent on the number of recorded neurons, with fewer neurons leading to less informative results.
  • While the model can capture temporal dynamics, the specific length of history modeled by cRBMs is an additional parameter that needs to be chosen.
  • Future Work:
    • The authors suggest further investigations into the applicability of RBM modeling for decoding large-scale spiking activity in other brain circuits.
  • They envision this approach supporting the development of future brain-machine interfaces for clinical applications.
  • Exploring how different RBM architectures or hyperparameters might further improve the modeling of complex neural dynamics is a natural extension.

Summary updated

Abstract

The retina is a complex circuit of the central nervous system whose aim is to encode visual stimuli prior the higher order processing performed in the visual cortex. Due to the importance of its role, modeling the retina to advance in interpreting its spiking activity output is a well studied problem. In particular, it has been shown that latent variable models can be used to model the joint distribution of Retinal Ganglion Cells (RGCs). In this work, we validate the applicability of Restricted Boltzmann Machines to model the spiking activity responses of a large a population of RGCs recorded with high-resolution electrode arrays. In particular, we show that latent variables can encode modes in the RGC activity distribution that are closely related to the visual stimuli. In contrast to previous work, we further validate our findings by comparing results associated with recordings from retinas under normal and altered encoding conditions obtained by pharmacological manipulation. In these conditions, we observe that the model reflects well-known physiological behaviors of the retina. Finally, we show that we can also discover temporal patterns, associated with distinct dynamics of the stimuli.
BibTeX
@inproceedings{volpi2020modeling,
  author = {Riccardo Volpi and Matteo Zanotto and Alessandro Maccione and Stefano Di Marco and Luca Berdondini and Diego Sona and Vittorio Murino},
  title = {{Modeling a population of retinal ganglion cells with restricted Boltzmann machines}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-73691-z},
}
Figure