Back to the abstracts page
Back to the NCCD 2015 home page
Dynamics of evidence integration in canonical models of perceptual decision making
Genis Prat Ortega1 2, Klaus Wimmer2 Jaime de la Rocha2 and Alex Roxin2
1IDIBAPS 2Centre de Recerca Matemàtica

The brain is able to interpret streams of high-dimensional ambiguous information and yield coherent percepts. This process of evidence integration has been studied extensively [1], but the underlying neuronal network dynamics remain largely unknown. Several models can explain the behavioral performance in these tasks [1,2], although they rely on different dynamical mechanisms.

Here, we characterize the dynamics of evidence integration using three canonical one-dimensional models: (1) the Drift Diffusion Model (DDM), (2) the Perfect Integrator (PI) and (3) the double-well potential (DW) which captures the dynamics of the attractor networks [3]. We examined the categorization dynamics of these models in response to fluctuating stimuli of different duration seeking contrasting and experimentally testable predictions. Stimuli are drawn from a Gaussian distribution N(μ, σ) and the two stimulus categories are defined by μ > 0 and μ ≤ 0. We find that the models behave differently in response to stimuli with different σ: In the small σ regime, both the DW and the DDM perform transient integration and exhibit a decaying psycophysical kernel (PK) revealing a primacy effect (i.e. only the first part of the stimulus is used for classification). In the large σ regime, the effective integration window in the DDM decreases because the absorbing bound is reached earlier. In contrast, the DW allows for classification reversals whose rate increases with σ producing a change in the PK time course from decreasing to increasing (recency effect). The PI shows constant PK for all σ's (i.e. uniform integration). The discrimination sensitivity (β) in the PI and the DDM decreases monotonically with σ because the two stimulus categories become less separable and because of the reduction in effective integration time in the DDM. Remarkably, the β in the DW as a function of σ shows a local maximum at σmax > 0. This "stochastic resonance" phenomenon allows some correcting reversals ( i.e. transitions to the deeper well) to compensate for the decrease in stimulus separability. Finally, we also find that the models behave differently with σ when the initial condition is offset, what can represent unequal prior probabilities of the two categories: the choice bias caused by the offset decreases with σ in the DW, increases in the DDM and remains constant in the PI.

Our analysis makes strong specific predictions to be tested in psychophysical two alternative forced-choice tasks which would pin down the basic principles of sensory integration dynamics.

[1] Gold, Shadlen. J. Annu. Rev. Neurosci. 30:535-574 (2007).
[2] Wang. Neuron 36:955-968 (2002).
[3] Roxin, Ledberg PLoS Comput Biol 4:9e1000046 (2008)