SieveColorSpace

V4 fMRI calibration — results summary

Dataset

• Source: OpenNeuro ds005521 (Conway et al., J. Neurosci. 2025)
• License: CC0 (public domain)
• Subjects: 2 macaques (sub-M1, sub-M2)
• Runs: 118 fMRI runs
• Design: 9 DKL hues × 4 contrasts = 36 conditions, ~1062 stimulus blocks of 14 s at TR = 2 s
• Pipeline: functional ROI (top 10% activated, posterior cortex), block-average BOLD, HRF-shifted +4 s, percent signal change

Key numerical results (cited in PT_COLOR.tex abstract and §Relation to DKL)

V4 channel weights at 95% contrast

Channel	V4 weight (BOLD)	SCS prediction (γ_p / Σγ_p)	Gap
L − M	0.373	0.385	3.2% ← headline match
S	0.056	0.284	80% (S-cone underrepresented in functional ROI)
Luminance	0.571	0.750	24%

The L − M match is the claim cited in the paper. The S and Luminance gaps are honest caveats: the functional ROI underrepresents S-cone signal (well-known fMRI limitation) and the luminance channel carries contributions from non-chromatic visual processing that the SCS simplex does not separate.

PT hypothesis tests

Test	Result	Verdict
Luminance (p=2) > Chromatic	BOLD ratio 2.57	Confirmed
L − M (γ₃) > S (γ₇)	L − M −0.078% vs S −0.086%	Marginal, same sign
r(BOLD, d_Fisher), 8 chromatic hues	0.434	Positive, not dominant
r(BOLD, d_Fisher) at 30% contrast (linear regime)	0.517	Best correlation

V4 contrast-response decomposition

Contrast	L − M (γ₃)	S (γ₇)	Luminance (p=2)	|LM/S|
10%	−0.154	−0.041	+0.206	3.75
30%	−0.095	−0.115	+0.079	0.83
50%	−0.055	−0.042	+0.043	1.31
95%	−0.137	+0.021	+0.210	6.62

SCS predicted |LM/S| = γ₃ / γ₇ = 1.356; observed mean 3.13. The agreement is qualitative across contrasts, not a tight quantitative match.

What the V4 data confirms

• The luminance/chromaticity split (p = 2 vs {3, 5, 7}) is echoed in V4 cortical responses (luminance 2.57× stronger than chromatic).
• The L − M channel weight in V4 matches γ₃ / (γ₃ + γ₅ + γ₇) to 3.2% — a prediction from s = 1/2 with 0 adjusted parameters.
• Low-contrast V4 responses correlate with Fisher distance on the SCS simplex (r = 0.517 at 30% contrast).

What the V4 data does NOT do

• It does not replace CIECAM02 in the hybrid model — the V4-opponent hybrid reaches r = 0.675 on COMBVD, vs r = 0.824 for SCS + CIECAM02 with actual CAM02 features.
• The S channel is underrepresented in our functional ROI — a known fMRI limitation, not a theoretical failure.
• 9 DKL directions cannot interpolate to arbitrary COMBVD colors without additional modeling.

What the V4 data suggests as follow-up

• Proper retinotopic mapping (task-meridianmapper) for a true V4 ROI instead of the functional proxy used here.
• GLM analysis with motion correction for cleaner BOLD estimates.
• Denser DKL stimulus grid to drive out V4 tuning curves.
• Replication on human fMRI data (not just macaque).

Reproducibility

python3 scripts/v4_neural_extraction.py      # pipeline
python3 scripts/v4_refined_analysis.py       # prints L-M = 0.373 vs γ₃/Σγ = 0.385
python3 scripts/v4_hybrid_model.py           # COMBVD hybrid with V4 opponent channels
python3 scripts/v4_analysis_plots.py         # figures

Data file paths are resolved via the SCS_V4_DATA environment variable; default is data/ds005521 under the repo root. Download instructions are in the v4_neural_extraction.py header.

Cited files

• datasets/v4_bold_response.csv — 36 conditions (9 hues × 4 contrasts)
• datasets/v4_neural_transfer.csv — model comparison table
• docs/figures/fig_v4_bold_polar.pdf
• docs/figures/fig_v4_bold_vs_fisher.pdf
• docs/figures/fig_v4_contrast_response.pdf
• docs/figures/fig_v4_gamma_ratio_test.pdf
• docs/figures/fig_v4_simplex_map.pdf

This site is open source. Improve this page.