Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Approximate Bayesian computation with a complex internal model naturally combines probabilistic inference and heuristics

Ádám Koblinger¹, József Fiser¹; ¹Académie d'Aix-Marseille

Presenter: Ádám Koblinger

Proposals differ on how the brain accounts for the uncertainty of perceptual variables – either by representing them as probability distributions that explicitly encode uncertainty in their width (Knill & Pouget, 2004), or by exploiting the correlation between the uncertainty of one variable (e.g., orientation) and the value of others (e.g., contrast), using the latter’s point estimates as heuristic proxies (Bertana, Chetverikov, van Bergen, Ling, & Jehee, 2021). The two approaches offer distinct advantages – probabilistic representations provide superior data- and memory-efficiency, while proxy-based strategies impose substantially lower computational demands – and each has its proponents, depending on which advantage is considered more relevant to brain function (Barthelme & Mamassian, 2010; Meyniel, Sigman, & Mainen, 2015; Koblinger, Fiser, & Lengyel, 2021). Rather than strictly contrasting these hypotheses, we follow a normative perspective and argue that both strategies can emerge naturally in a unified framework when time-evolving approximate inference is optimized to solve realistic tasks involving the joint estimation of multiple interacting variables. We formalize this idea by modeling behavior as the output of an ideal observer that combines approximate probabilistic perceptual representations with fast, coarse proxy information – yielding a flexible hybrid approach. Through simulations, we show that the model adaptively relies on proxies to compensate for the coarseness of approximate inference. Finally, by directly comparing the model’s output to empirical data, we demonstrate that observed behavior qualitatively aligns with the predictions of this hybrid model.

Topic Area: Predictive Processing & Cognitive Control

Extended Abstract: Full Text PDF