Amplitude modulations of sensory responses, and deviations from Weber’s Law in pulsatile evidence accumulation

Sue Ann Koay, Stephan Y. Thiberge, Carlos D. Brody, David W. TankbioRxiv 2020 · DOI

Abstract

How do animals make behavioral decisions based on noisy sensory signals, which are moreover a tiny fraction of ongoing activity in the brain? Some theories suggest that sensory responses should be accumulated through time to reduce noise. Others suggest that feedback-based gain control of sensory responses allow small signals to be selectively amplified to drive behavior. We recorded from neuronal populations across posterior cortex as mice performed a decision-making task based on accumulating randomly timed pulses of visual evidence. Here we focus on a subset of neurons, with putative sensory responses that were time-locked to each pulse. These neurons exhibited a variety of amplitude (gain-like) modulations, notably by choice and accumulated evidence. These neural data inspired a hypothetical accumulation circuit with a multiplicative feedback-loop architecture, which parsimoniously explains deviations in perceptual discrimination from Weber-Fechner Law. Our neural observations thus led to a model that synthesizes both accumulation and feedback hypotheses.

Introduction: Beyond Feedforward Processing

In classical neuroscience, sensory processing is often conceptualized as a linear, feedforward pipeline. Information flows from sensory organs to the primary cortex, then to higher-order areas where it is integrated to inform a decision. However, this model struggles to account for the pervasive feedback loops found throughout the brain and the specific behavioral nuances of evidence accumulation.

A key focus of this research is Weber’s Law, which posits that the "just-noticeable difference" between two stimuli is proportional to their magnitude. While this law is a cornerstone of psychophysics, recent studies in rodents have shown a curious deviation: at low levels of evidence, performance improves more rapidly than Weber’s Law predicts. This study investigates the neural mechanisms behind this phenomenon, proposing that sensory neurons are not passive relays but active participants in a feedback-modulated decision process.

Methods and Experimental Design

To explore these dynamics, researchers at Princeton University employed a sophisticated experimental setup combining virtual reality (VR) and high-resolution imaging.

Figure 1: The 'Accumulating-Towers' task design and behavioral performance. (a-c) Mice navigate a VR T-maze with random visual cues (towers). (d) Two-photon imaging map of posterior cortex. (e-g) Behavioral data showing psychometric curves and performance across evidence levels.

The study utilized two-photon calcium imaging to record the activity of over 10,000 neurons across six areas of the posterior cortex, including the primary visual cortex (V1). Mice performed the "Accumulating-Towers" task, navigating a virtual corridor where visual cues (towers) appeared randomly on either side. The mice were rewarded for turning toward the side with the higher tower count. This task requires the animal to maintain a running tally of evidence while ignoring noise and distractors.

Key Insights: Context-Aware Sensory Neurons

The researchers identified a subset of "cue-locked" neurons that fired in direct response to each visual tower. Surprisingly, these responses were not uniform.

Figure 2: Neural response modeling and regional activity. (a-b) Calcium traces for individual cells during right and left choice trials. (c-e) The impulse response model used to characterize cue-locked activity. (f-j) Distribution and properties of cue-locked cells across different cortical layers and regions.

Analysis revealed that even in V1, approximately 50% of these cue-locked cells were modulated by the animal's current evidence or its eventual choice. This suggests that "higher-order" decision variables are fed back to the earliest stages of visual processing. The researchers found that this feedback acts as a multiplicative gain. For instance, if an animal is biased toward a right turn, the neural response to right-side cues is amplified.

Reconciling Weber's Law

The most significant theoretical contribution of this work is the "multiplicative feedback-loop" model ($fdbk$). This model explains the transition from sub-Weber to Weber-like performance:

  • Low Evidence: At low cue counts, the system is limited by basic sensory noise. As more cues are added, performance improves quickly because the signal-to-noise ratio increases.
  • High Evidence: At high counts, the noise within the feedback loop itself begins to scale with the signal. This modulatory noise eventually dominates, leading to the constant-ratio performance characteristic of Weber’s Law.

The $fdbk$ model significantly outperformed traditional feedforward and simple integrator models in explaining behavioral data from both mice and rats, providing a unified framework for understanding evidence accumulation.

Limitations and Future Directions

While the study provides a compelling model for feedback-driven decision-making, several questions remain. The physical location of the "accumulator"—the circuit responsible for the feedback signal—is still unknown. Potential candidates include the parietal cortex, frontal eye fields, or subcortical structures. Additionally, the exact biological implementation of multiplicative gain (e.g., through specific inhibitory interneuron types) requires further investigation.

Future work utilizing optogenetics to selectively disrupt these feedback loops could provide causal evidence for this model, potentially shifting animal behavior from a Weber-like regime back to a purely feedforward state.

Summary updated

Abstract

Abstract How do animals make behavioral decisions based on noisy sensory signals, which are moreover a tiny fraction of ongoing activity in the brain? Some theories suggest that sensory responses should be accumulated through time to reduce noise. Others suggest that feedback-based gain control of sensory responses allow small signals to be selectively amplified to drive behavior. We recorded from neuronal populations across posterior cortex as mice performed a decision-making task based on accumulating randomly timed pulses of visual evidence. Here we focus on a subset of neurons, with putative sensory responses that were time-locked to each pulse. These neurons exhibited a variety of amplitude (gain-like) modulations, notably by choice and accumulated evidence. These neural data inspired a hypothetical accumulation circuit with a multiplicative feedback-loop architecture, which parsimoniously explains deviations in perceptual discrimination from Weber-Fechner Law. Our neural observations thus led to a model that synthesizes both accumulation and feedback hypotheses.
BibTeX
@misc{koay2020amplitude,
  author = {Sue Ann Koay and Stephan Y. Thiberge and Carlos D. Brody and David W. Tank},
  title = {{Amplitude modulations of sensory responses, and deviations from Weber’s Law in pulsatile evidence accumulation}},
  year = {2020},
  doi = {10.1101/2020.06.24.167213},
  howpublished = {bioRxiv},
}
Figure