Marius Usher and James L. McClelland (2001), Psychological Review 108(3):550–592 — DOI-10.1037/0033-295X.108.3.550
Key takeaways
- A choice among N alternatives can be modelled as N noisy accumulators. Each one leaks, each one inhibits the others, and the first to reach an absolute criterion wins. Lateral inhibition makes each unit’s activation reflect relative evidence, so the absolute criterion of accumulator models and the relative-evidence character of the diffusion model are both kept, and the scheme extends to any number of alternatives without change.
- For two alternatives, linearised, the difference x₁ − x₂ is an Ornstein–Uhlenbeck process whose net leak is K = k − β (unitwise leak minus inhibition). When leak and inhibition balance (K = 0) the process is the classical diffusion model, and accuracy grows without bound. When they do not, asymptotic d′ is finite, 2v/σ·√(1/K) (Eq. 10). The time–accuracy curve depends only on |K|, so dominant leak and dominant inhibition give the same accuracy over time but different trajectories: leak-dominated systems weight late evidence (recency) and inhibition-dominated ones weight early evidence (primacy).
- Leakage, not only trial-to-trial drift variance (Ratcliff 1978), can explain why accuracy levels off, and the two predict differently shaped time–accuracy curves. The empirical support is modest: on 3 participants the leaky (OU) curve beats the drift-variance curve in 7 of 9 conditions, with a combined likelihood ratio of about 62. The authors themselves conclude that both factors probably operate.
Standing in the record
Filed on 2026-10-03 at the owner’s request, with the diffusion decision model (LIT-576), the optimality analysis that unifies the two (LIT-569), Schurger, Sitt and Dehaene’s use of a leaky accumulator for self-initiated movement (LIT-605), and the HDDM estimation package (LIT-563). These five are the record’s first statement of evidence-accumulation models of decision and action initiation. No anthology topic holds a model of human perceptual choice, and the paper carries no instruction for machine-learning practice.
It is the source of the competing-accumulator half of the family. Bogacz et al. (LIT-569) take its “mutual inhibition” model and show that, balanced and with large leak and inhibition, it collapses onto the diffusion model and is then optimal. That is the formal version of a point this paper made by simulation. Schurger et al. (LIT-605) take a single leaky accumulator from it, cited by them as “a leaky stochastic accumulator process”, and use it to model the decision of when to move. Ratcliff and McKoon (LIT-576) treat it as one of the newer accumulator models that “may be as successful as” the single-process diffusion model, while noting they have been tested on fewer paradigms.
For the record’s unities, it bears on behavioral-integration in a narrow and
precise sense: how many candidate responses come down to one through
competition. That is the decision-level analogue of action selection, not a
claim about how an organism’s whole control hierarchy integrates. The record’s
nearest existing account of competing motive states is Frijda et al. on
emotion regulation (LIT-540, THEORY-056). Lateral inhibition between
accumulators is a formal model of one state’s readiness suppressing another’s,
but nothing in this paper is about emotion or motive, and no relation is
claimed.
No relation is declared. The paper compares itself against Ratcliff’s 1978 diffusion-with-drift-variance model, which the record does not hold, and the Ratcliff and McKoon review in the record postdates it.