Magnitude estimation reveals Poisson-like noise underlying perception

This study demonstrates that analyzing magnitude estimation ratings reveals a single stochastic representation of visual contrast comprising a sigmoidal transducer and Poisson-like noise, which quantitatively predicts discrimination sensitivity and explains classic perceptual phenomena without free parameters.

The Gist

Imagine your brain is a super-advanced radio receiver trying to tune into a station (a visual stimulus, like a striped pattern). The goal of this research was to figure out exactly how that radio works: How does it turn a physical signal into what you feel you are seeing? And more importantly, why does the radio sometimes get fuzzy or static?

For decades, scientists tried to understand this by playing two different songs and asking, "Which one is louder?" (a discrimination task). But this approach had a blind spot. It was like trying to figure out if a car engine is loud because the engine is powerful or because the exhaust pipe is broken, without being able to see inside the hood. You could measure the noise, but you couldn't separate the engine's power from the static.

The New Approach: The "Volume Knob" Experiment

The researchers in this paper tried a different trick. Instead of just asking "Which is louder?", they asked participants to rate how loud the sound was on a scale of their own choosing. This is called Magnitude Estimation.

Think of it like this:

• Old Way (Discrimination): "Is this cup of coffee hotter than that one?" (Yes/No).
• New Way (Magnitude Estimation): "On a scale of 1 to 100, how hot is this coffee?"

The genius of this study is that they didn't just look at the average rating (e.g., "It feels like a 50"). They looked at the variability of the ratings. If you ask someone to rate the same coffee ten times, do they say "50, 50, 50" every time? Or do they say "48, 52, 49, 55"?

That "wobble" or "jitter" in the answers is the key. It represents the internal noise in the brain—the static on the radio line.

What They Discovered: The Brain's "Sigmoid" Shape

By analyzing both the average ratings and the "wobble," the researchers found two main things about how our brains process visual contrast (how distinct stripes look against a background):

1. The "S-Curve" Transducer (The Volume Knob)
The brain doesn't turn up the volume linearly. It uses a special S-shaped curve (a sigmoid).

• At very low volumes (faint images): The brain is super sensitive. It turns the volume up fast to make sure you don't miss a whisper. This is like a "pedestal effect"—a little bit of signal makes a huge difference in perception.
• At high volumes (bright images): The brain compresses the signal. It turns the volume up more slowly so you don't get overwhelmed. This is why doubling the brightness of a bright light doesn't feel like it's twice as bright.

2. The "Poisson" Noise (The Static)
The researchers found that the "static" or noise in the brain isn't random background hiss. It's signal-dependent.

• Analogy: Imagine a microphone. When the room is quiet, the background hiss is low. But as you speak louder, the microphone picks up more hiss proportional to your voice.
• In the brain, the louder the signal, the more "noise" there is. This is called Poisson-like noise. It's like the universe saying, "The more you try to see, the fuzzier the edges get."

The Big Reveal: One Brain, Two Tasks

The most exciting part of the paper is that they proved one single model explains both ways of testing the brain.

They took the data from the "Rate the contrast" task (Magnitude Estimation), calculated the brain's internal "S-curve" and "noise level," and then used a math formula to predict how well people would do on the "Which is louder?" task (Discrimination).

The prediction was perfect.
Without changing any numbers, the model derived from the subjective ratings accurately predicted the objective discrimination results. It successfully recreated two famous, confusing phenomena:

• The Pedestal Effect: Why we are better at spotting a faint signal if there's a tiny bit of background noise (the "S-curve" helps us).
• Weber's Law: Why it gets harder to tell the difference between two bright lights than two dim ones (the "noise" gets louder as the signal gets stronger).

The Takeaway

This study is like finding the instruction manual for the human brain's visual system. It tells us that:

1. Subjective feelings (ratings) and objective performance (tests) come from the same place.
2. The brain is a smart filter: It amplifies weak signals to catch them, but accepts that strong signals will always come with some fuzziness.
3. We can measure the invisible: By simply asking people to rate how strong a sensation feels, we can mathematically map out the hidden "noise" and "wiring" inside their minds.

In short, the brain isn't a perfect camera; it's a clever, slightly noisy radio that uses a specific set of rules to turn the world into what we see. And now, we know exactly what those rules are.

Technical Summary

1. Problem Statement

The fundamental challenge in sensory neuroscience is determining how physical stimuli map to perceptual experience. Traditionally, researchers rely on discrimination sensitivity ($D(x)$), defined as the ability to detect an increment ($\Delta x$) over a baseline intensity ($x$). However, sensitivity measurements alone suffer from an identifiability problem:

• Sensitivity is determined by the signal-to-noise ratio (SNR), which is the slope of the transducer function ($\mu(x)$) divided by the internal noise ($\sigma(x)$).
• Equation: $D(x) = \frac{|\mu'(x)|}{\sigma(x)}$.
• The Ambiguity: Different combinations of transducer gain and noise levels can yield identical sensitivity curves. Therefore, discrimination data alone cannot uniquely identify the internal noise distribution or the specific form of the transducer function.

While recent studies have attempted to resolve this by combining sensitivity with reaction times or subjective ratings, direct empirical evidence linking subjective rating variability to internal noise remains limited.

2. Methodology

The authors employed a dual-task experimental design with 11 participants using visual contrast (Gabor patches) as the stimulus.

A. Experimental Tasks

1. Magnitude Estimation (ME):
   • Participants viewed gratings with contrasts ranging from 0.05% to 14%.
   • They provided unconstrained verbal numerical ratings of perceived intensity.
   • Hypothesis: The mean of these ratings reflects the transducer function $\mu(x)$, and the variability (standard deviation) of these ratings directly reflects the internal noise $\sigma(x)$.
2. Two-Interval Forced-Choice (2IFC) Discrimination:
   • Participants compared a reference stimulus and a comparison stimulus to judge which had higher contrast.
   • This provided empirical discrimination sensitivity profiles ($D_{discr}$) to validate the predictions derived from the ME task.

B. Data Analysis & Modeling

• ME Modeling: The response distributions were modeled as a normal distribution $N(\mu(x), \sigma(x)^2)$. The authors tested various functional forms for the transducer and noise.
  • Transducer Model: A non-saturating sigmoidal function (Eq. 2) capturing both expansive (low contrast) and compressive (high contrast) nonlinearities.
  • Noise Model: A power-law noise function with an additive baseline (Eq. 3): $\sigma(x) = \sigma_0 + q \cdot \mu(x)^\gamma$.
• Prediction: Using the parameters derived solely from ME (mean and variance), the authors calculated predicted discrimination sensitivity using the standard SNR equation ($D_{pred} = \frac{|\mu'(x)|}{\sigma(x)}$) without any free parameters for the discrimination task.
• Validation: The predicted sensitivity ($D_{pred}$) was compared against the empirically measured discrimination sensitivity ($D_{discr}$) derived from psychometric functions fitted to the 2IFC data.

3. Key Contributions

1. Direct Access to Internal Noise: The study demonstrates that the variability in magnitude estimation ratings is not just "response noise" but a direct behavioral proxy for the brain's internal stochastic representation.
2. Unification of Psychophysical Frameworks: It provides empirical evidence unifying Stevens' Power Law (subjective magnitude) and Fechner's Law (discrimination thresholds). It shows that a single internal representation (transducer + noise) underlies both subjective ratings and objective discrimination.
3. Mechanistic Decomposition: The study mathematically disentangles the sources of sensitivity limits, attributing the pedestal effect to transducer nonlinearity and Weber-like behavior to signal-dependent noise.

4. Key Results

A. Characterization of Internal Representation

• Transducer Function: The data revealed a sigmoidal transducer.
  • Low Contrast: Exhibits an expansive nonlinearity ($\mu(x) \propto x^\beta$) with a steep exponent ($\beta \approx 4.68$).
  • High Contrast: Exhibits a compressive nonlinearity ($\mu(x) \propto x^\alpha$) with an exponent $\alpha \approx 0.74$.
• Noise Structure: The internal noise follows a signal-dependent power law ($\sigma(x) \propto \mu(x)^\gamma$) with an exponent $\gamma \approx 0.64$.
  • This is Poisson-like (where variance scales with the mean, implying $\gamma=0.5$) but slightly supra-Poisson ($\gamma > 0.5$).
  • A small additive noise term ($\sigma_0$) was present but numerically small for most participants.

B. Prediction of Discrimination Phenomena

Using only the ME-derived parameters, the model successfully predicted the two hallmark phenomena of contrast discrimination without free parameters:

1. The Pedestal Effect: At very low contrasts, sensitivity increases with intensity. The model attributes this to the expansive transducer ($\beta \approx 4.68$) amplifying the signal faster than the noise increases.
2. Weber-like Behavior: At higher contrasts, sensitivity decreases as intensity increases ($D(x) \propto x^{-1}$). The model attributes this to signal-dependent noise overtaking the compressive transducer gain.

C. Quantitative Agreement

• The predicted sensitivity curves ($D_{pred}$) closely matched the empirical discrimination curves ($D_{discr}$) in shape, peak position, and Weber exponents.
• While the absolute magnitude of predicted sensitivity was slightly lower than observed (likely due to known biases in ME scaling), the functional form and individual differences (specifically peak position and Weber exponents) were highly correlated between the two tasks.

5. Significance and Implications

• Resolving the Identifiability Problem: By utilizing the variability of magnitude estimation, the study breaks the degeneracy between transducer gain and noise level, allowing for a unique reconstruction of the internal representation.
• Biological Plausibility:
  • The derived sigmoidal transducer aligns with contrast response functions observed in the primary visual cortex (V1) of animals, human fMRI, and EEG.
  • The supra-Poisson noise mirrors the signal-dependent variability found in cortical neurons.
• Theoretical Synthesis: The findings suggest that the "limits" of perception are not caused by a single mechanism. Instead:
  • Low Contrast: Limits are defined by the transducer's expansive nonlinearity (thresholding).
  • High Contrast: Limits are defined by signal-dependent (Poisson-like) noise.
• Methodological Shift: The paper argues that the variability of subjective ratings, historically overlooked, is a critical metric for understanding the stochastic nature of perception, offering a more parsimonious explanation for sensory processing than models relying solely on discrimination thresholds.