TL;DR
This paper investigates how the brain uses sensory information to make decisions, specifically focusing on "choice probabilities" (CPs) – the correlation between a neuron's activity and a subject's decision. By using biologically-constrained simulations of visual and vestibular heading perception, the authors demonstrate that while correlated neural noise is necessary for CPs to emerge, selective decoding of neural signals plays a crucial role in shaping these CPs and matching behavioral data.
What Problem Does This Paper Address?
When we make decisions based on sensory input, our brains don't just process the "average" signal from neurons. There's also "noise" – random fluctuations in neural activity – and how this noise is structured, along with how the brain "reads out" or decodes this activity, can influence our choices. A key measure here is "choice probability" (CP), which quantifies how often a neuron's firing pattern predicts a specific decision. For decades, there's been a debate: do CPs primarily reflect the inherent "noise" in neural populations, or do they reveal how the brain selectively uses specific neural signals for decision-making? This paper aims to disentangle these two factors, particularly in the context of multisensory integration for heading perception.
Methods
The authors employed a multi-pronged approach:
- Biologically-Constrained Simulations: They built computational models of neural populations that mimic the properties of neurons in the extrastriate visual cortex (specifically area MSTd) that process self-motion (heading) information from both visual (optic flow) and vestibular (inner ear) cues.
- Two Core Models:
- Pure-Correlation Model: This model posits that CPs arise solely from correlated noise between neurons, where some neurons contribute to the decision and others don't, but their noise is linked.
- Selective Decoding Model: This model assumes that neurons contribute to the decision, and their signals are selectively "read out" or weighted by a decoder based on their tuning properties.
- Multisensory Integration: They specifically focused on "opposite cells" – neurons that have conflicting preferences for visual and vestibular heading cues. This unique property provided a powerful test case.
- Comparison with Experimental Data: The models' predictions for noise correlations, choice probabilities, and behavioral performance were directly compared against existing experimental data from MSTd neurons and animal behavior.
Datasets
The study utilized existing datasets from previous experiments on MSTd neurons in monkeys performing heading discrimination tasks. Specifically, they used data from 129 single neurons and behavioral data from two monkeys. For their simulations, they generated population responses based on the statistical properties (tuning curves, noise correlations) derived from these real neural datasets.
Key Insights and Ideas
- Noise is Necessary, but Not Sufficient: While correlated noise is essential for CPs to be observable in large neural populations, it alone cannot fully explain the observed patterns, especially in multisensory contexts.
- Selective Decoding is Key for Multisensory CPs: The peculiar pattern of CPs in "opposite cells" (positive CPs for vestibular cues, negative for visual cues) is better explained by a model where the brain selectively decodes these neurons based on their vestibular heading preference, rather than just relying on correlated noise.
- Decoding Strategy Matters: The way the brain "reads out" neural information (the decoding strategy) significantly shapes CPs, particularly when dealing with conflicting sensory inputs.
- Vestibular Preference as a Decoding Anchor: The data strongly suggest that the brain might preferentially use vestibular heading information to decode signals from both congruent and opposite cells, providing a robust way to estimate heading.
Key Results
- Noise Correlation Structure: The selective decoding model's assumption about noise correlations (depending on both visual and vestibular tuning similarity) provided a significantly better fit to experimental data than the pure-correlation model's assumption (depending only on vestibular tuning).
- Choice Probabilities: Both models could replicate the basic pattern of CPs for congruent and opposite cells in single-cue conditions. However, in the combined (multisensory) condition, the selective decoding model predicted a unimodal distribution of CPs for opposite cells, which matched experimental data, whereas the pure-correlation model predicted a bimodal distribution, which did not.
- Behavioral Performance: The selective decoding model, with a specific weighting of opposite cells, accurately predicted the animals' psychophysical performance (thresholds) across vestibular, visual, and combined conditions, aligning with optimal cue integration theory. The pure-correlation model failed to match these behavioral predictions.
Limitations and Future Work
- Model Simplifications: The models used simplified representations of neurons and decoding strategies. For instance, the decoding was assumed to be optimal and to have knowledge of the correlation structure, which might not hold true in biological systems.
- Focus on MSTd: The study focused on neurons in area MSTd. Future work could explore whether these findings generalize to other brain areas involved in decision-making.
- Further Disentangling Factors: While the study strongly favors selective decoding, the authors acknowledge that both correlated noise and selective decoding likely contribute to CPs in a complex interplay. Future research could further refine the understanding of how these factors interact.
- Robustness of Decoding: The paper suggests that decoding based on vestibular preference might offer robustness to moving objects in optic flow. Further investigation into the functional advantages of this specific decoding strategy is warranted.
Summary updated