Storyboard

Object, event, and risk recognition is the problem of assigning a semantic category (prey, predator, companion, obstacle) to sensory patterns. Bayesian decision theory provides the optimal framework: maximizing expected utility by weighting the posterior probability of each hypothesis by the consequences of each decision.

The Mahalanobis distance normalizes the distance to the centroid of each class by the variability of that class (¹): it avoids distortion due to correlations between features. In Gaussian linear classification, the optimal decision boundary is the hypersurface where d²_M is equal between two classes, which is a hyperplane (linear discriminant analysis).

Signal detection theory (SDT) separates sensitivity (d', which depends on SNR) from response criterion (, which depends on a priori probabilities and consequences). An observer can have high sensitivity (high d') and still make errors if his criterion is biased. Applying TDS to animal behavior allows sensory limits to be separated from decision biases.

Diffusion evidence integration (race model) accumulates sensory evidence up to a threshold _dec: it models reaction time and error rate simultaneously. The drift velocity _drift is proportional to the SNR of the stimulus; the threshold _dec controls the speed-accuracy trade-off. Neurons in IPL (inferior parietal lobe) of primates show ramping activity consistent with this model.

The perceived risk function R_perc introduces cost asymmetry: in dangerous situations, C_damage >> C_falsepos, which makes the action threshold very low (better to flee from a false alarm than not to flee from a real predator). This evolutionary asymmetry explains the exaggerated fear responses and negativity biases observed in animals.

>Model

ID:('ky', 585)

Description

ID:(0, 585)