Neuron's eye view: Inferring features of complex stimuli from neural responses

Xin Chen, Jeffrey M. Beck, John M. PearsonPLOS Computational Biology 2017 · OpenReview

TL;DR

This paper introduces a novel Bayesian model that can automatically discover underlying features of complex stimuli directly from neural responses, without prior knowledge of those features. This approach is crucial for understanding how the brain processes complex, unstructured information, such as natural scenes or social interactions, by moving beyond traditional methods that require pre-defined feature sets.

What Problem Does This Paper Address?

Understanding how the brain represents information from the world is a fundamental goal in neuroscience. Traditionally, researchers study this by presenting stimuli with carefully controlled features (like contrast or pitch) and observing how individual neurons respond. However, this approach has a major limitation: it requires experimenters to already know which features are important. This is problematic for complex domains like social interactions or natural movement, where the relevant features are poorly understood. The authors aim to overcome this by developing a method that can discover these hidden stimulus features solely from neural data, thus avoiding confirmation bias and enabling exploration of more complex stimuli.

Methods

The core of this paper is a Bayesian model that treats neural responses as being driven by multiple, discrete, time-varying latent features of the stimulus. Key aspects of their approach include:

  • Latent Feature Discovery: The model assumes that neural firing rates are modulated by a set of unknown latent features that evolve over time with Markovian dynamics.
  • Bayesian Inference: They employ a fully variational Bayesian approach to infer both the model parameters (e.g., how strongly neurons respond to each feature) and the latent stimulus features themselves. This allows for principled model comparison and provides uncertainty estimates.
  • Hierarchical Priors: Sparse hierarchical priors are used on the feature modulation strengths. This encourages parsimony, ensuring that only features strongly supported by the data are inferred, effectively implementing Occam's razor.
  • Fast Inference Algorithm: They derive a fast variational Bayesian inference algorithm that uses coordinate ascent updates and forward-backward inference for latent states, making it computationally efficient.

Datasets

The authors evaluate their model on two types of datasets:

  • Synthetic Data: Generated from their own model to demonstrate its ability to recover ground-truth features.
  • Labeled Neural Data:
    • Roitman Dataset: Single neuron recordings from macaque area LIP during a perceptual discrimination task involving randomly moving dots. This dataset tests the model's ability to infer features from unlabeled data, even when units are recorded independently.
  • Visual Category Dataset: Recordings from macaque inferotemporal cortex neurons presented with images from 8 categories (including faces, bodies, and scenes). This dataset allows for comparison of inferred features to known stimulus categories.

Key Insights and Ideas

  • Unsupervised Feature Discovery from Neural Data: The most significant insight is the demonstration that complex stimulus features can be reliably inferred from neural population activity without any prior knowledge of those features. This opens the door to studying brain responses to highly complex, naturalistic stimuli.
  • Parsimony through Hierarchical Priors: The use of sparse hierarchical priors is crucial for preventing the model from inferring spurious features. It effectively forces the model to only identify features that have a substantial impact on neural responses across the population.
  • Latent Features as "Tags": The model conceptualizes latent features as time-varying "tags" of the stimulus, which can be combined in a combinatorial way to represent complex stimuli. This provides a flexible framework for capturing rich stimulus information.
  • Model's Ability to Handle Ambiguity: The model can correctly identify features even when the data is noisy or when there are ambiguities in the stimulus labels, as seen in the Roitman dataset where it correctly identified fewer features than were present in the experimental design.

Key Results

  • Synthetic Data: The model successfully recovered the ground-truth latent features from synthetic data, demonstrating its core functionality. It also showed robustness to inferring more features than were present, by assigning high uncertainty to the unused ones.
  • Roitman Dataset: The model inferred 5 features from a dataset where 10 were designed, accurately capturing the most significant experimental variables (e.g., target location relative to receptive field).
  • Visual Category Dataset: The model successfully identified features corresponding to categories like monkey faces and bodies, and showed that these features could be combined additively to represent more complex stimuli (e.g., close-ups vs. general monkey images). The inferred features also captured subtle distinctions, like the "dead time" at the beginning of stimulus presentation.
  • Computational Efficiency: The variational Bayesian inference algorithm is described as "exceptionally fast," allowing for efficient analysis.

Limitations and Future Work

  • Computational Complexity for Semi-Markov Models: While the general framework is efficient, inference for semi-Markov duration distributions requires explicit optimization (BFGS), which can be computationally intensive for a large number of states.
  • Assumption of Independent Latent Chains: The current model assumes that each latent feature corresponds to an independent Markov chain. Future work could explore dependencies between these latent features.
  • Potential for More Efficient Optimization: The authors note that more efficient optimization routines could likely be developed for some of the non-conjugate updates.
  • Future Directions: The authors suggest that this approach is well-suited for exploring complex domains like social neuroscience and olfaction, where hypotheses about neural representations are scarce. They envision an iterative experimental process where stimuli are first presented in a large, heterogeneous dataset, then tagged based on neural responses, and finally refined to accurately capture neural correlates.

Summary updated

BibTeX
@article{chen2017neuron,
  author = {Xin Chen and Jeffrey M. Beck and John M. Pearson},
  title = {{Neuron's eye view: Inferring features of complex stimuli from neural responses}},
  journal = {PLOS Computational Biology},
  year = {2017},
  url = {https://openreview.net/forum?id=jDNf5I4BAd},
}
Figure