PaperlayerView exact revision on arXiv

Source-linked AI summary

Structured Bayesian Modeling of Dynamic Receptive1 Fields in Salamander Retinal Ganglion Cells

Alokesh Manna

arXiv:2609.30731v1cs.NEstat.ME

TL;DR

Estimating dynamic receptive fields from few trials requires handling many correlated pixel-by-time coefficients, while ordinary sparse pixel selection does not enforce spatial coherence or temporal smoothness. The paper fits a Bayesian Poisson model with spatial and temporal structure to each of 155 neurons and clusters their temporal profiles. The fits yield coherent fields rather than fragmented Poisson-LASSO patterns, while clustering raw surfaces gives a degenerate grouping.

  • Problem

    Receptive-field estimation from few trials involves many correlated pixel-by-time predictors, while ordinary LASSO does not encode spatial organization.

    Source: §1 · pp. 2–3

  • Method

    The paper fits a Poisson model with a spatial Gaussian Markov random field and temporal AR(1) process independently for each neuron, then clusters temporal response profiles.

    Source: §4 · p. 11, §10 · p. 37

  • Results

    The per-neuron fits recover coherent surfaces rather than fragmented Poisson-LASSO patterns, while clustering raw surfaces gives a degenerate grouping.

    Source: §8 · p. 34, §10 · p. 37

  • Takeaways & Limitations

    The structured Bayesian formulation provides a coherent representation of dynamic receptive-field estimation, with the joint population fit remaining a target for future work.

    Source: §5.2 · p. 24, §10 · p. 37

    Population typing follows independent per-neuron fits and post hoc clustering, so it does not pool information within phenotypes or demonstrate the gain from joint fitting.

    Source: §5.2 · p. 24

Abstract

from arXiv · show

Neurons in the visual system are selective for specific spatial and temporal stimulus features, described by their \emph{receptive field}. Estimating one means a coefficient per pixel per time bin from few trials -- a high-dimensional problem requiring regularization. Sparse regularizers such as the LASSO handle the dimension but select pixels independently at each time point, with nothing to keep the region coherent in space or smooth in time; it can fragment or reorganize discontinuously even when the true response evolves smoothly, a failure since this evolving pattern is what a receptive-field estimate should capture. We formulate dynamic receptive-field estimation as a high-dimensional Bayesian problem: a Poisson model combining a Gaussian Markov random field in space with an autoregressive process in time, so the estimated field is smooth and coherent across space and time. On recordings from $155$ salamander retinal ganglion cells, fitting this model independently per neuron recovers a coherent surface, where a pixel-level Poisson-LASSO comparison instead returns a fragmented one. Summarizing each neuron's surface by its space-averaged temporal response and clustering these curves with a model-based functional-clustering procedure, BIC selects three balanced temporal-response phenotypes ($85$, $32$, $38$ neurons), against a degenerate grouping from clustering the raw surfaces. A simulation study with known ground truth confirms the same pattern, with the model beating an unregularized Poisson GLM, LASSO, and the elastic net on recovery and estimation accuracy, though LASSO controls false positives better. The per-neuron field identification, its contrast with LASSO, and the functional-clustering population typing constitute this paper's contribution.

1 Introduction

The paper frames receptive-field estimation as recovering which image locations influence firing and how those effects evolve over time. It proposes a structured Bayesian approach to address high dimensionality and spatial dependence while representing neuron heterogeneity and shared type structure.

1 Introduction

  • With few trials, 169 pixels across 30 time bins create 5,070 candidate coefficients per neuron, while neighboring pixels are correlated and useful estimates should identify coherent regions.Across 155 neurons, a separate coefficient for every pixel and time bin would yield about 786,000 parameters, making standard MCMC impractical at this scale.Source: §1 · pp. 2–3
  • Ordinary LASSO can select pixels independently despite their spatial dependence, so it does not ensure a spatially organized receptive-field estimate.The paper identifies questions about spatial location, temporal evolution, shared features, and neuron-specific deviations, alongside whether structured sparsity can improve recovery over LASSO.Source: §1 · pp. 3–4
  • The proposed formulation treats receptive fields as structured spatio-temporal objects and aims to borrow information across neurons of the same functional type.The hierarchical model is intended to address spatial and temporal dependence, sparsity, neuron heterogeneity, and shared structure across types.Source: §1 · p. 4
  • The fitted results use a spatial GMRF and temporal autoregression independently per neuron; the shared type-level component and structured-sparsity extension are specified but not jointly fitted across the population.The paper combines established SPDE, AR(1), and INLA methods into a structured representation that makes fitting tractable at the dataset’s scale.Source: §1 · pp. 4–5

2 Neural Spike Data

The study uses repeated salamander retinal ganglion cell recordings of spike counts evoked by grayscale images. The images are coarsened to a 13 × 13 grid for analysis.

2 Neural Spike Data

  • 303 grayscale images were each presented 13 times, producing 3,939 trials with spike counts recorded in 30 bins over each 200-millisecond trial.The original 256 × 256 images were coarsened to 13 × 13 pixels.Source: §2 · p. 6
  • Figure 1 shows four representative 13 × 13 stimulus images, with pixel intensities mean-centered within each image.This mean-centered intensity scale is used as the covariate in the model fits.Source: §2 · p. 6, Figure 1 · p. 7

3 Motivation: Why LASSO Is Not Sufficient

The section illustrates why sparse pixel selection alone is insufficient for receptive-field analysis: it neither encodes spatial coherence nor addresses correlated predictors, effect shrinkage, temporal dependence, or neuron heterogeneity.

3.1 Ordinary sparse regression

  • The L1 penalty encourages zero coefficients but assigns the same penalty to coherent and spatially fragmented configurations with equal numbers of nonzero pixels.The two example configurations each contain four nonzero coefficients, though only one forms a coherent region.Source: §3.1 · pp. 7–8

3.2 A motivating empirical illustration

  • At fixed illustrative λ = 0.02 on full pooled data, Poisson-LASSO yielded 27 of 169 nonzero pixels, including isolated selections beyond a central cluster; the image-level baseline showed a middle band but could not localize a receptive field.Source: §3.2 · pp. 8–9, Figure 3 · p. 9

3.3 Correlated image predictors

  • Neighboring pixels can carry similar information, and the scientific target is a receptive-field region rather than necessarily an individual pixel.The elastic net encourages correlated variables to enter or leave together, unlike LASSO’s tendency to select only one from a correlated group.Source: §3.3 · p. 10

3.4 Shrinkage bias

  • Because LASSO shrinks nonzero coefficients toward zero, it may distort coefficient magnitudes that carry information about response strength.The same penalty performs both sparsity selection and estimation.Source: §3.4 · p. 10

3.5 Spatial and temporal dependence

  • Receptive-field coefficients are expected to relate across neighboring pixels and adjacent time bins, dependencies ordinary LASSO does not explicitly encode.Prior work cited in the section reports more accurate and interpretable estimates from spatial-locality priors when data are limited relative to stimulus dimensions.Source: §3.5 · pp. 10–11

3.6 Heterogeneous neurons

  • Neurons may belong to different functional types, so fitting each independently ignores potential information sharing while forcing one field on all neurons ignores heterogeneity.The motivation is a hierarchical model that allows both sharing and differentiation.Source: §3.6 · p. 11

4 The Hierarchical Structured-Sparsity Model

The fitted model treats each neuron’s receptive-field surface as a latent Gaussian field, combining spatial structure and temporal dependence in a Poisson model. INLA is used to fit this high-dimensional model, while pooling across neurons is only proposed as future work.

4.2 Part A — the model we fit

  • Each neuron is fit independently with a latent receptive-field surface, a baseline intercept, repeated-presentation random effects, spatial GMRF structure, and AR(1) dependence across time bins.The model uses an SPDE representation for the spatial field; its sparse precision matrix couples neighboring mesh nodes, discouraging isolated selections.Source: §4 · p. 11, §4.2 · pp. 12–14

4.1 Neural response model and observation family

  • The Poisson observation model fits nonnegative spike counts over fixed time windows, but its equal-mean-and-variance assumption was not formally tested against residuals.A negative-binomial extension is listed as follow-up work.Source: §4.1 · p. 12

4.3 Bayesian inference and the role of INLA

  • INLA replaces MCMC sampling of the high-dimensional latent field with a Laplace approximation and integration over four hyperparameters, exploiting the sparse precision matrix.The model is a latent Gaussian model, and the paper describes MCMC as impractical within a reasonable time frame at this data scale.Source: §4.3 · pp. 16–17
  • Pooling information across neurons of the same functional type is sketched as a future extension, but it has not been fit and contributes no reported results.The results in the paper use the per-neuron model.Source: §4 · p. 11, §4.3 · p. 18

5 Real-Data Analysis

The real-data analysis fits a structured spatio-temporal model to one salamander neuron and then independently fits all 155 neurons to identify temporal-response phenotypes. The population analysis finds three clusters, while the joint hierarchical model and whether phenotype-level pooling adds value remain future work.

5.1 Single-neuron demonstration

  • The fitted single-neuron surface highlights a central image region, where coefficients fluctuate more during time window 16–20.This is a demonstration on real data with an unknown receptive field; Figure 5 shows the spatio-temporal coefficients.Source: §5 · p. 18, §5.1 · p. 20, Figure 5 · p. 20

5.2 Population-level neuron typing

  • For all 155 neurons, the analysis independently fits spatio-temporal surfaces, then clusters each surface’s space-averaged temporal profile.This two-stage procedure does not impose shared phenotype-level structure during fitting.Source: §5.2 · p. 21
  • BIC-selected functional clustering yields three balanced temporal-response phenotypes with 85, 32, and 38 neurons.The clusters are defined by fitted response-curve shape, not established biological cell classes.Source: §5.2 · pp. 21–22, Figure 6 · p. 23
  • All three phenotypes show an early positive response, a mid-trial decline, and partial late recovery, differing in decline timing and depth.Phenotype 2 declines earliest and most sharply; phenotype 1 declines latest and most gradually, while phenotype 3 recovers to the highest final value.Source: §5.2 · p. 22, Figure 6 · p. 23
  • The phenotypes’ magnitude-weighted centroids lie near the grid center, while their peak pixels lie near different peripheral locations.The peak is a single-pixel summary, so the spatial differences are suggestive rather than conclusive; the analysis has not tested correspondence to established retinal ganglion cell classes.Source: §5.2 · p. 23, Figure 7 · p. 24
  • The independent-fit clustering does not pool information within phenotypes or show what such pooling would add.The authors propose fitting the joint model with G = 3 and quantifying the gain as future work.Source: §5.2 · p. 24

6 Interpreting the fitted coefficients

The fitted coefficient surface has a direct interpretation in terms of stimulus-associated excitation or inhibition. INLA’s marginal posteriors also provide neuron-level probabilities of positive coefficients; corresponding type-level quantities belong to an extension not yet fit.

6 Interpreting the fitted coefficients

  • A positive β_kst indicates an excitatory association between pixel s and neuron k’s firing at time t; a negative value indicates an inhibitory association.Source: §6 · p. 24
  • INLA’s marginal posteriors provide the neuron-level probability P(β_kst > 0 | Y) at each pixel and time bin.The passage describes this as posterior evidence for excitation at that location.Source: §6 · p. 25
  • Type-level fields and membership probabilities are properties of the not-yet-fit extension, not outputs of the current Part A fit.Source: §6 · p. 25

7 Simulation Study

The simulation evaluates receptive-field recovery against known ground truth, using a computationally reduced slice of a broader design. Across the executed comparisons, spatio-temporal SPDE/AR(1) leads on predictive score and coefficient error, while false-positive control is mixed.

7.1 Design

  • The executed design is a reduced slice: two spatial configurations, G ∈ {1, 3}, and τδ ∈ {0.4, 1.6}, rather than the broader five-configuration, three-type, continuous-deviation design.Each combination used 12 neurons on a 13 × 13 grid and six time bins, reduced from 30 for computational tractability.Source: §7.1 · p. 26

7.2 Ground-truth receptive fields: exactly which pixels are active, and why

  • True fields use Gaussian bumps hard-zeroed beyond 2.5σ, making the active mask exactly the nonzero coefficients rather than a post-hoc threshold.The cutoff gives the simulation known inactive pixels for support-recovery scoring.Source: §7.2 · p. 27

7.3 Neurons and data generation

  • Individual deviations were implemented as localized, signed hard-cutoff bumps, not the target fine-grained GMRF draws; the authors identify literal GMRF deviations as remaining work.Synthetic observations used four repeated presentations per image, with one held out and three used for training.Source: §7.3 · p. 28

7.4 Methods and evaluation

  • The comparison included GLM, LASSO, elastic net, spatial SPDE, and SPDE/AR(1); the proposed hierarchical structured-sparsity model was not implemented or run.The SPDE methods were fit only for the first four neurons per combination; formal component-count comparison, temporal stability, and uncertainty calibration remain future work.Source: §7.4 · pp. 29–30

7.5 Results

  • Correcting the ground truth to exact zeros lowered TPR for every method (mean absolute difference 0.06) and changed which method appeared best on that criterion.Log-score and coefficient error changed little between runs, with mean absolute differences of 0.02 and 0.03, respectively.Source: §7.5 · p. 30
  • 8/8 combinations: SPDE/AR(1) led log-score (mean −1.93 vs GLM −2.68, LASSO −2.74) and coefficient error (mean 0.99 vs LASSO 1.42, GLM 1.54).Table 1 covers compact fields and Table 2 disconnected fields; spatial and spatio-temporal SPDE results average the first four neurons per combination.Source: Table 1 · p. 31, Table 2 · p. 31, §7.5 · p. 33
  • Mean TPR was 0.869 for SPDE/AR(1), 0.859 for GLM, and 0.650 for LASSO; LASSO had the lowest mean FPR at 0.152 versus 0.276 for SPDE/AR(1).No method dominated TPR by combination: GLM and spatial SPDE led in three combinations each, and SPDE/AR(1) in two.Source: §7.5 · p. 33
  • Across combinations, GLM and both SPDE variants beat the GLM?Source: §7.5 · pp. 33–34

8 Discussion

The paper frames receptive-field estimation as a structured problem, but its fitted analyses do not yet implement the proposed joint sharing across neuron types. The real-data phenotypes remain unvalidated as biological types, and the analysis is limited to one experimental preparation.

8 Discussion

  • Clustering raw 5,070-dimensional coefficient surfaces produced an outlier-dominated split, unlike the usable population typing from space-averaged temporal profiles.The temporal-profile representation’s robustness to alternative summaries, clustering methods, selection criteria, and SPDE hyperparameters has not been tested.Source: §8 · pp. 34–35
  • The hierarchical joint model is sketched but not fully specified or fit; reported results use independent per-neuron fits followed by post hoc clustering.Pooling within neuron types and its benefit over independent fits therefore remain unmeasured.Source: §8 · p. 35
  • The three temporal-response phenotypes are empirical partitions, not validated biological classifications, because they were not compared with an independently established classification.The paper also leaves the robustness of the representation and clustering choices untested.Source: §8 · p. 35
  • The paper reports no quantitative held-out comparison against simpler alternatives on real data; real neurons lack known true active sets for assessing support recovery.The simulation study, not a real-data predictive comparison, supplies the primary quantitative evidence about recovery and coefficient error.Source: §8 · p. 35
  • The three-phenotype finding comes from one salamander retinal ganglion cell preparation and stimulus set, so its generality remains unestablished.Independent recordings or another sensory system would be needed to assess whether the structure extends beyond this dataset.Source: §8 · p. 36
  • Further directions include a negative-binomial likelihood, neural-history terms, flexible spatial priors, theoretical recovery analysis, and testing transfer to human visual-cortex mapping.The proposed extension to population receptive-field mapping in human visual cortex is a future question, not a demonstrated result.Source: §8 · p. 36

9 Future Work

Future work is to fit a joint model that shares receptive-field information among neurons of the same functional type while retaining neuron-specific deviations. This model remains a target rather than a completed or evaluated method.

9 Future Work

  • The proposed next step is to pool information across neurons of the same functional type within the model, rather than only after independent field estimation.Each field would decompose into a shared type-level component and a neuron-specific deviation.Source: §9 · p. 36
  • The type-level field would be smoothed in space and time and given a sparsity mechanism to avoid reintroducing fragmented support.This design is proposed for the shared field, not reported as a fitted model.Source: §9 · p. 36
  • The joint extension is future work: it is not fully specified or fit, and no paper result depends on it.The authors identify completing it as the highest-priority next step.Source: §9 · p. 37

10 Conclusion

The paper demonstrates independent per-neuron spatio-temporal receptive-field fits, contrasts their coherent surfaces with fragmented Poisson-LASSO estimates, and clusters temporal profiles into three phenotypes. Its proposed joint hierarchical extension remains future work.

10 Conclusion

  • On 155 salamander retinal ganglion cells, independent SPDE/AR(1) fits produce coherent receptive-field regions, while pixel-level Poisson-LASSO returns fragmented patterns.In simulations with known ground truth, the spatio-temporal model outperforms an unregularized Poisson GLM, LASSO, and elastic net on recovery and coefficient accuracy, while LASSO controls false positives better.Source: §10 · p. 37
  • Clustering the fitted temporal response profiles yields three balanced phenotypes, with counts of 85, 32, and 38 neurons.This is a post hoc population typing based on the independent per-neuron fits.Source: §10 · p. 37
  • The demonstrated results use independent per-neuron fits; the joint type-sharing extension has not been fully specified or fit.Its contribution over independent fits has therefore not been measured.Source: §10 · p. 37
  • The approach reframes receptive-field estimation as structured high-dimensional Bayesian inference, with improved recovery and interpretability across heterogeneous populations stated as potential once the joint fit is complete.The conclusion presents this broader potential conditionally, not as an outcome established by the current joint model.Source: §10 · pp. 37–38

11 Data and Code Availability

The analysis code is available from the corresponding author upon request, with a public repository release planned alongside the completed simulation study.

11 Data and Code Availability

  • Analysis code for the LASSO comparison, per-neuron INLA/SPDE fits, and population clustering is available from the corresponding author upon request.A public repository release is planned alongside the completed simulation study, with figure-generating scripts intended to support replication.Source: §11 · p. 38

Appendix A Population-level clustering: additional detail

The appendix details how temporal-response phenotypes are derived from each neuron's fitted surface and how those labels are used to construct Figure 7. BIC selects three phenotypes with 85, 32, and 38 neurons.

Per-neuron temporal profile

  • Each neuron’s surface is averaged over its 169 pixels at each of 30 time bins to form a temporal profile.This yields 155 curves, each a 30-point vector.Source: Appendix A · p. 40

B-spline expansion

  • Each 30-point curve is least-squares projected onto an 8-function cubic B-spline basis, producing a smoothed, lower-dimensional representation.The resulting vectors have length 8.Source: Appendix A · p. 40

Functional PCA

  • PCA is applied to the 155 × 8 coefficient matrix, retaining at least two components until cumulative explained variance reaches 99%.Source: Appendix A · p. 40

Gaussian mixture and BIC

  • A Gaussian mixture is fitted to the retained PCA scores for candidate cluster counts and covariance families, and BIC chooses the best combination.Source: Appendix A · p. 40
  • BIC selects k = 3, with cluster sizes 85, 32, and 38 neurons.Each neuron receives one of three phenotype labels.Source: Appendix A · p. 40

Cluster-average spatio-temporal field and peak

  • For each phenotype, member neurons’ full fitted surfaces are averaged into a 13 × 13 × 30 field; the largest-absolute-value cell identifies its plotted triangle.The absolute value treats excitatory and inhibitory coefficients alike when locating the strongest response.Source: Appendix A · p. 41

Magnitude-weighted centroid

  • The plotted circle is a magnitude-weighted centroid of all 169 pixels at the peak’s time bin, with every pixel included and no threshold.Source: Appendix A · p. 41

the role of its hyperparameters

The spatial precision matrix couples neighboring mesh nodes and leaves non-neighboring pairs uncoupled. Its hyperparameters separately govern spatial range and overall marginal variance.

Spatial coupling

  • Qw is sparse because its mesh-adjacency structure gives zero entries to non-neighboring nodes, making it a graph-Laplacian-type operator.The construction follows the SPDE approach and uses a finite-element discretization of a differential operator on the mesh.Source: the role of its hyperparameters · p. 41
  • The GMRF conditional mean pulls each node toward a precision-weighted average of immediate neighbors, unlike the pixel-independent LASSO penalty.Source: the role of its hyperparameters · p. 42

Hyperparameters

  • κ controls spatial range and relative neighbor coupling, while σw rescales marginal variance without changing that coupling pattern.Both hyperparameters receive penalized-complexity priors, leaving the data to determine the degree of spatial coupling.Source: the role of its hyperparameters · p. 42
Loading 2609.30731v1…