Feed 0% source
Mathematics AI-generated

Numerosity adaptation reflects multiple levels of sensory processing

Generated by a local model (nvidia/Gemma-4-26B-A4B-NVFP4) from a scientific paper, claim-checked against the full text. Provenance is open by design.

When we look at a crowd of people or a scattering of dots, our brains perform a rapid, subconscious estimation of how many items are present. This ability—known as numerosity perception—is a fundamental skill. It underlies everything from basic survival instincts to complex mathematical reasoning. Most scientists believe this "sense of number" is handled by specialized, high-level neurons in the brain's parietal cortex. These neurons are thought to be indifferent to the specific appearance of the objects they count.

However, recent evidence suggests that this math-oriented part of the brain might be easily distracted by low-level visual details. When we stare at a large group of white dots, our subsequent ability to estimate numbers changes. This is called numerosity adaptation. The big question is whether this change happens because we have "exhausted" our number-counting neurons. Or is it simply a side effect of our eyes getting tired of seeing specific colors or contrasts?

A new study from the University of Osaka and the University of Nottingham proposes that the answer is both. The researchers report that numerosity adaptation is not a single-step process. Instead, it is a two-stage mechanism involving both early sensory processing and high-level numerical representation.

The breakdown of the pure number hypothesis

The prevailing view in neuroscience is that numerosity perception is robust to low-level visual features. These include density, area, or contrast. Under this "direct effect" model, the brain extracts a pure number from a scene. It ignores the "how" of the visual input. If this were true, adapting to a large number of black dots should produce the same numerical aftereffect as adapting to white dots. The "number" being processed would be identical.

Current research has hit a snag. Some studies found that changing the color or contrast between the adaptation and test phases weakens the numerical aftereffect. This suggests the "pure number" model is incomplete. If the brain focused only on quantity, the visual "flavor" of the dots should not matter. The existing gap is whether these visual features interfere directly with math neurons. Or do they merely muddy the signal before it reaches them?

A two-stage processing architecture

To resolve this, the authors propose a hierarchical model of visual processing. Think of it like a digital camera. The sensor (early sensory stage) captures light and color to create a raw signal. The image processor (high-level stage) then interprets that signal to identify objects and quantities.

The authors' mechanism works in two distinct steps:

  1. Low-level sensory adaptation: At the earliest stages of vision, the brain adapts to physical properties. One example is contrast polarity (whether a dot is lighter or darker than the background). This is like how your eyes adjust to a bright room before entering a dark one.
  2. High-level numerosity adaptation: Once the signal moves up the chain, specialized neurons tune themselves to quantity. These neurons are meant to be "feature-invariant." This means they should ideally only care about the count, not the color.

The logic is that if the first stage is fatigued by a specific contrast, it sends a biased signal to the second stage. The authors argue that visual features change our number perception because the "raw data" is altered by early-stage sensory exhaustion.

Evidence from mismatch and masking

The researchers tested this by mismatching the contrast of the dots. In Experiment 1, they found that "mixed-polarity" adaptors (equal parts black and white dots) caused different perceived numbers than single-color adaptors .

Figure 1
Figure 1. Stimulus sequence and results of Experiment 1. (A) In each trial, adaptor and probe stimuli were presented sequentially, each consisting of two dot arrays displayed in the left and right visual fields. During the response period (red fixation), participants responded which of the two dot arrays in the probe stimulus contained more dots. In the mixed-polarity condition, the target adaptor consisted of an equal number of black and white dots, whereas in the single-polarity condition all dots shared the same contrast polarity (either black or white), which matched that of the probe stimulus. For illustration purposes, fewer dots are shown than

Specifically, the Point of Subjective Equality (PSE)—the threshold where a person perceives two different stimuli as equal—shifted significantly [Figure 1C]. This shift shows that perceived numerosity is sensitive to contrast polarity.

The most critical evidence came from Experiment 2. Here, the authors manipulated "congruency" (whether the dots in the adaptation phase matched the test phase). They found that the numerical aftereffect was reduced when colors did not match (incongruent conditions). However, it was never completely eliminated [Figure 2C]. This is a vital distinction. If the effect were purely a low-level visual trick, the mismatch should have destroyed the effect entirely. The persistent aftereffect proves a high-level "number" mechanism is still active.

To rule out retinal afterimages (the "ghost images" seen after staring at bright lights), the authors performed Experiment 3. They used a white-noise mask to "wipe" the visual field. They also reduced contrast to minimize lingering images .

Figure 4
Figure 4. Stimulus sequence and results of Experiment 3. (A) Except for the insertion of a 500-ms mask stimulus immediately following the adaptor and probe stimuli, the stimulus sequence was identical to that of Experiment 2. (B) Psychometric functions fitted to the mean proportion of 'test greater' responses across participants. Solid and dashed lines indicate congruent and incongruent contrast polarity conditions, respectively. Blue, red, and black lines represent the small (20 dots), large (80 dots), and neutral (40 dots) adaptation conditions. (C) Mean numerosity aftereffect magnitudes for congruent and incongruent contrast polarity conditions across participants. Gray circles represent individual participants' data. Error bars denote standard error. ** p < 0.01; *** p < 0.001.

Even with these controls, robust numerical aftereffects remained [Figure 4C]. They also found no correlation between reported afterimage intensity and the strength of numerical errors .

Figure 5
Figure 5. Correlations between numerosity aftereffect magnitude and Likert scale scores of afterimages in the incongruent condition. Correlation between aftereffect magnitude and rated intensity (A) and frequency (B) of afterimages following the adaptor stimulus. Correlation between aftereffect magnitude and rated intensity (C) and frequency (D) of afterimages following the probe stimulus. Each dot represents an individual participant. Spearman's rank correlation coefficients (ρ) are shown in each panel.

Limits of the two-stage model

While the study provides a compelling framework, it is not exhaustive. The authors admit several limitations.

First, the study focused only on contrast polarity. It is unknown if this mechanism applies to other dimensions like color, shape, or motion. Second, the researchers used a fixed range of numerosities. Human vision handles small groups (subitizing) differently than large textures. It is unclear if this dual-layer adaptation holds true across those scales. Finally, the study does not distinguish between different spatial reference frames. This leaves a question regarding whether these two stages of processing occur in the same "coordinate system" of the brain.

The verdict: A hybrid reality

The evidence supports a hybrid verdict. Numerosity perception is neither a purely high-level calculation nor a simple low-level reflex. It is an integrated process. Early sensory filters shape the data that high-level processors eventually decode.

The authors' computational model successfully reproduced these behavioral patterns by combining both layers . This suggests the "sense of number" is a layered construction. For researchers, the takeaway is practical. Future studies should carefully control for contrast polarity. Failing to do so may lead to misinterpreting how much of a result comes from pure number processing versus low-level sensory interference. To understand how an observer perceives "how many," you must also account for the "how" of the visual signal.

Figures from the paper

Figure 2
Figure 2. Stimulus sequence and results of Experiment 2. (A) As in Experiment 1, adaptor and probe stimuli were presented sequentially. The target adaptor contained either 20 (small), 40 (neutral), or 80 (large) dots, while the numerosity of the remaining stimuli (i.e., the neutral adaptor, test stimulus, and reference stimuli) was identical to that in Experiment 1. During the
Figure 3
Figure 3. Aftereffect magnitude for black and white probes. Filled circles connected by a solid line represent the mean aftereffect magnitude in the congruent and incongruent conditions for the black-probe group (n = 8), whereas open circles connected by a dashed line represent the corresponding mean for the white-probe group (n = 8). Gray lines show individual participants' data. Error bars denote standard error.
Figure 6
Figure 6. Results of Experiment 4. (A) Psychometric functions fitted to the mean proportion of 'test greater' responses across participants. Solid and dashed lines indicate congruent and incongruent contrast polarity conditions, respectively. Blue, red, and black lines represent the small (20 dots), large (80 dots), and neutral (40 dots) adaptation conditions. (B) Mean numerosity aftereffect magnitudes for congruent and incongruent contrast polarity conditions across participants. Gray circles represent individual participants' data. Error bars denote standard error. *** p < 0.001.
Novelty
0.0/10
Overall
0.0/10
#research
How this was made
Generation

Model: nvidia/Gemma-4-26B-A4B-NVFP4
Persona: academic_accessible
Template: engineering_deepdive
Refinement: 0
Pipeline: forge-1.1

Verification

Evaluator: nvidia/Gemma-4-26B-A4B-NVFP4
Score: 93% (passed)
Claims verified: 17 / 18

Translation

Model: nvidia/Gemma-4-26B-A4B-NVFP4

Hardware & cost

NVIDIA GB10 · 128 GB unified · NVFP4 · 100% local · $0 cloud
Tokens: 119,682
Wall-time: 264.0s
Tokens/s: 453.3