TL;DR
This study investigates how primate area MT, crucial for motion perception, represents both visual stimuli and behavioral choices. Using advanced dimensionality reduction on neural population data, they found that choice information is largely encoded in a subspace orthogonal to the stimulus representation. This suggests a "non-corrupting feedback" mechanism, where downstream decision circuits inform sensory areas without degrading the original stimulus signal, potentially aiding learning and adaptation.
What Problem Does This Paper Address?
When animals make decisions based on sensory information, especially near perceptual thresholds, the activity of sensory neurons often correlates with the animal's reported choice. A common interpretation is that downstream decision-making circuits "read out" this neural activity, and any variability in the sensory neurons directly influences the choice. However, this "feedforward" view is complicated by the possibility of feedback from decision areas to sensory areas. The challenge is to disentangle whether choice-correlated activity in sensory areas is a consequence of how downstream circuits read the information (feedforward) or a result of top-down influence (feedback). This paper aims to resolve this by examining the population-level representation of both stimulus and choice in primate area MT.
Methods
The researchers employed a variational latent Gaussian process (vLGP) method to extract low-dimensional latent factors from simultaneously recorded MT neuron activity. This technique is novel in its ability to handle spike train data and capture shared trial-to-trial variability without requiring trial-averaging, thus preserving information about choice.
Key steps included:
- Latent Factor Extraction: Using vLGP to identify underlying shared dynamics in the population activity.
- Subspace Alignment: Rotating these latent factors to align one dimension with the stimulus direction ("stimulus axis") and the remaining dimensions with the "non-stimulus subspace."
- Choice Mapping: Developing a "choice mapping" technique to project multivariate neural activity onto a one-dimensional representation that reflects behavioral choice. This allowed for the calculation of choice probability (CP) along different dimensions (stimulus axis, non-stimulus subspace, and the whole population).
- Nested Likelihood Ratio Tests: Comparing different models (e.g., stimulus-only, stimulus + non-stimulus, stimulus + non-stimulus + population) to determine the contribution of each dimension to predicting choice.
- Time Course Analysis: Analyzing the temporal evolution of choice probability to distinguish between feedforward and feedback hypotheses.
Datasets
The study used electrophysiological data recorded from three adult rhesus macaque monkeys performing a motion discrimination task. The data consisted of spike trains from simultaneously recorded MT neurons across multiple sessions. The paper does not introduce new benchmark datasets.
Key Insights and Ideas
- Stimulus Information is Highly Localized: The visual motion stimulus was primarily encoded in a single latent dimension (the "stimulus axis") across the MT population, consistent with MT's role as a sensory area.
- Choice Information is Misaligned: Surprisingly, a significant portion of choice-related information was found in the latent subspace orthogonal to the stimulus axis. This "misalignment" directly challenges the simplest feedforward readout models.
- Non-Corrupting Feedback Hypothesis: The findings strongly support a "non-corrupting feedback" hypothesis. This suggests that downstream decision circuits send signals back to MT, but these signals are channeled into the non-stimulus subspace, allowing them to influence choice without degrading the stimulus representation.
- Temporal Dynamics Support Feedback: The time course of choice probability in the non-stimulus subspace showed a slow rise, peaking later in the trial. This temporal profile is more consistent with feedback mechanisms accumulating over time rather than immediate feedforward readout.
- Maintaining Mixed Representations: This misaligned feedback mechanism allows MT to maintain a mixed representation of sensory input and decision-related signals, which could be crucial for learning and adapting to new tasks or changing environments.
Key Results
- The first latent dimension explained over 90% of the stimulus-related variability in the MT population.
- Choice probability in the stimulus axis was close to chance (e.g., 0.546 across sessions), indicating minimal direct influence of stimulus-aligned neural variability on choice.
- Choice probability in the non-stimulus subspace was significantly above chance (e.g., 0.591 across sessions), demonstrating that choice information resides in dimensions orthogonal to the stimulus.
- The pooled choice probability across all latent dimensions was 0.621, indicating that the combined representation carries substantial choice information.
- Time course analysis showed that choice probability in the non-stimulus subspace increased gradually and peaked later in the trial, supporting a feedback interpretation.
Limitations and Future Work
- Linear Readout Assumption: While the vLGP method captures nonlinear dynamics, the analysis of choice mapping and model comparison relied on logistic regression, which assumes a linear readout of the latent factors. The authors acknowledge that optimal nonlinear readout could alter interpretations.
- Generalizability: The study focused on primate area MT. Future work could explore whether this "misaligned feedback" mechanism is conserved in other sensory areas or species.
- Functional Role of Feedback: While the study suggests feedback is non-corrupting and potentially aids learning, further experiments are needed to directly test its role in adaptation and task learning. The authors hypothesize that this feedback might be useful for tuning receptive fields and learning optimal readouts in relation to task context.
Summary updated