Story
Prior conditioning shifts the psychometric, not the slope
A textbook prediction in perceptual decision-making: telling a subject
that one option is more likely should bias them toward it (
μ
moves) without changing their underlying sensitivity to the stimulus
(
σ
stays put). Two papers in the atlas test exactly this — once in mice
via block-level prior schedules, once in humans via trial-by-trial
cues. They use different mechanics, but they make the same
prediction. The atlas's canonical psychometric axis lets us read
both off side by side.
Mouse
IBL 2021
trainingChoiceWorld biased blocks, p(left) ∈ {0.2, 0.5, 0.8}, the prior is set per block (~90 trials each). The reference here is the unbiased 0.5 block; deltas report the bias and slope change relative to it.
| Condition | n | μ (bias) | σ (slope) | Δμ | Δσ |
|---|---|---|---|---|---|
| p_left=0.2 | 199 | -25.24 | 12.22 | -12.90 | 3.49 |
| p_left=0.5 (reference) | 90 | -12.34 | 8.73 | — | — |
| p_left=0.8 | 248 | -3.13 | 16.31 | 9.22 | 7.58 |
Human
Walsh 2024
Trial-by-trial cue (valid / neutral / invalid) of the upcoming target side. Cue is delivered at trial onset, ~75% predictive when valid. The reference here is the neutral cue; deltas report the bias and slope change relative to it.
| Condition | n | μ (bias) | σ (slope) | Δμ | Δσ |
|---|---|---|---|---|---|
| cue=invalid | 10,985 | -19.59 | 2.06 | -7.67 | -4.57 |
| cue=neutral (reference) | 10,960 | -11.92 | 6.63 | — | — |
| cue=valid | 44,255 | -1.73 | 1.19 | 10.20 | -5.44 |
The takeaway
Both species behave like the classical model predicts: μ moves substantially with the prior, and σ moves much less. The mouse biased-block effect is operationalised at the block timescale (priors persist for ~90 trials) and the human cue effect is operationalised at the trial timescale (cue resets every trial), but the canonical parameters land in the same place — μ shifts in the cued/biased direction, σ stays roughly fixed at each species' baseline.
The atlas's job here is to make this readable in one glance. The underlying findings are 4-parameter logistic fits run at build time on the harmonized canonical-trial CSVs from the IBL and Walsh slices; if you want to refit interactively, every curve is also available on /findings with the in-browser scipy fitter enabled.
Caveats
- IBL biased blocks and Walsh trial-cues operate at very different timescales. The atlas does not claim the underlying neural mechanism is the same — only that the behavioural read-out (psychometric μ vs σ) matches.
- Walsh's invalid-cue trials are rarer than valid; per-condition sample sizes differ. The slope estimate uncertainty is therefore not symmetric across conditions.
- Both fits use the 4-parameter logistic with γ and λ free. Bias-only is the textbook expectation but not the only possibility; a slope change would show up as a non-zero Δσ .