Errorless Irrationality: A unified computational account of the inverse base-rate effect across predictive, observational, and unsupervised procedures

The inverse base-rate effect (IBRE) is a robust bias in how people resolve ambiguity between competing categories, and the most prominent theories explain it through prediction error. Across two experiments we progressively removed the elements of the predictive-learning design that supply such error signals: first by moving to observational learning, then to an unsupervised procedure in which category labels were not presented. The effect persisted–the irrational bias is independent of supervised learning procedures. We propose a new theory, OSCAR, that integrates core computational principles of the best-validated models and operates on self-generated feedback akin to pattern completion. OSCAR extends the learning dynamics underlying the response bias to observational and unsupervised procedures. Evaluated on a large preexisting supervised dataset in addition to the two new experiments reported here, OSCAR performs competitively against alternatives, and is the first model that reproduces the pattern of individual differences seen in humans across all three procedures. The model provides an explanation of hitherto unexplained eye-tracking data, something none of the alternative accounts provide.

In Experiment 1, participants were presented with descriptions of people, the symptoms they experienced, and their respective diseases, with category labels presented simultaneously with features, removing explicit feedback. After exclusion, 125 participants made it into the main analysis. Participants exhibited a reliable common preference for A, MA = 0.68, 95% HDI [0.63, 0.73], BF10 = 2.45 × 10 7, and a reliable rare preference for BC, MBC = 0.67, 95% HDI [0.62, 0.72], BF10 = 1.11 × 10 7, confirming the IBRE in an observational procedure.

In Experiment 2, the procedure was implemented similar to a cued-recall task using solid black geometric shapes, removing any causal feature-label structure. After exclusion, 86 participants made it into the analysis. Participants showed a clear common preference for stimuli A, MA = 0.78, 95% HDI [0.73, 0.83], BF10 = 5.37 × 10 13, and a reliable rare preference on ambiguous BC trials, MBC = 0.73, 95% HDI [0.67, 0.79], BF10 = 8.12 × 10 8, further demonstrating the IBRE without design components that explicitly promote an error-driven process.

OSCAR (Outcome-Specific Configuration-dependent Attention Representation) is a feed-forward auto-associator network with two distinct layers. It uses an attention weight matrix with dimensional attention vectors, allowing features to acquire outcome-specific salience. The model combines attention weights with the stimulus vector to produce attention gain, which propagates through a competitive gating mechanism normalized row-wise. Output node activations are mapped to response probabilities using a gated softmax function. Error is calculated between the teaching signal and output activations, and attention is adjusted on gradient descent on error with respect to the underlying gains. The model uses a delta rule to update connection weights.

OSCAR produces the IBRE by combining outcome-specific attention with asymmetric weight structure: attention concentrates on the shared feature A and on the opposing-outcome unit, and reverses across the common and rare output rows for the perfect predictors B and C. The model learns equally strong inhibitory connections between B → rare and C → common, amplified through increased attention to those units, which counteract the excitatory connections formed between A and all other units. For BC, the model learns to inhibit common output unit for the AC feature configuration more strongly, driven by the strong base-rate following learned for A, than to inhibit the rare output unit for AB feature configurations, resulting in the BC → rare preference.

In model evaluation, OSCAR shares the win with NNRAS as the closest models to the PAS point with a g of 0.18 in the supervised paradigm. OSCAR accommodates slightly fewer patterns than NNRAS but scores more on reduced flexibility relative to NNRAS. g places DGCM in third place, and EXIT as the least adequate model. OSCAR and NNRAS remain the most adequate model for the widest range of weightings. For the observational and unsupervised evaluations, OSCAR's performance on accommodation within the zone of model adequacy, and its flexibility remains comparable across the paradigms.

OSCAR reproduces all key qualitative patterns in the data across all paradigms, including base-rate following for A and the irrational response bias for BC. The model also replicates the full set of fixation tendencies observed by Don et al. [2019], including the longer fixation for predictive features (B, C) over shared ones (A), the C over B advantage during training, and the B and C equivalence on BC trials, something no other model is able to accommodate.

OSCAR explains IBRE via a combination of attention, inhibition, and excitation. Two departures from the standard account underlie this: attention is a matrix of outcome-specific dimensional vectors rather than a single per-feature salience, and the network is an auto-associator in which features and category labels share an input space. Five consequences follow: goal-directed attention, salience modulated by feature configurations, feature absence being diagnostic, outcomes acquiring salience, and implications of uncertainty measures for confidence.

Improvements to AI Systems:

1. Add outcome-specific attention matrices instead of single per-feature salience weights. This lets models assign different attention to the same feature depending on the predicted outcome, enabling context-dependent feature weighting.

2. Implement auto-associative input spaces where features and labels share a common representation layer. This allows the system to treat absent features as informative (via pattern completion) and to propagate activation from labels back to features, mimicking self-generated feedback.

3. Replace error-driven supervised learning with pattern-completion-based learning for ambiguous or unsupervised settings. Use a delta rule on attention gains driven by internal reconstruction error (between input and auto-associator output) rather than external label error, enabling learning without explicit feedback.

4. Add competitive gating with row-wise normalization in the output layer. This creates asymmetric inhibition between outcome units, allowing the model to learn stronger inhibitory connections for rare vs. common configurations, producing irrational but human-like biases.

5. Introduce configuration-dependent salience modulation where attention gain is computed as a function of the full feature vector, not individual features. This lets the system dynamically reweight features based on co-occurring features (e.g., B and C together suppress common output more than rare).

6. Enable goal-directed attention by making attention weights outcome-specific, so the system can prioritize features that are diagnostic for a particular outcome while ignoring them for others, improving efficiency in multi-class tasks.

7. Model individual differences via flexibility-accuracy trade-off using a g-score (goodness-of-fit penalized by model flexibility). This allows AI systems to be tuned for robustness across different learning paradigms (supervised, observational, unsupervised) rather than overfitting to one.

What the Improved AI System Can Do:

• Learn category distinctions from purely observational data (no labels) or unsupervised exposure, matching human performance in ambiguous tasks.

• Exhibit human-like base-rate bias and irrational preference reversals when resolving ambiguous stimuli, useful for simulating human decision-making in cognitive modeling.

• Predict eye-tracking patterns (e.g., longer fixations on predictive features, asymmetric attention) without explicit training on gaze data, enabling more natural human-AI interaction.

• Handle missing or absent features as diagnostic signals, improving robustness in incomplete data scenarios.

• Adapt its attention strategy based on the outcome being predicted, leading to more flexible and context-aware classification in multi-label or hierarchical tasks.

• Provide a principled way to quantify model adequacy across different learning conditions, aiding in model selection for cognitive architectures.

The inverse base-rate effect is a robust bias in how people resolve ambiguity between competing categories, and the most prominent theories explain it through prediction error. Across two experiments we progressively removed the elements of the predictive-learning design that supply such error signals: first by moving to observational learning, then to an unsupervised procedure in which category labels were not presented. The effect persisted--the irrational bias is independent of supervised learning procedures. We propose a new theory, OSCAR, that integrates core computational principles of the best-validated models and operates on self-generated feedback akin to pattern completion. OSCAR extends the learning dynamics underlying the response bias to observational and unsupervised procedures. Evaluated on a large preexisting supervised dataset in addition to the two new experiments reported here, OSCAR performs competitively against alternatives, and is the first model that reproduces the pattern of individual differences seen in humans across all three procedures. The model provides an explanation of hitherto unexplained eye-tracking data, something none of the alternative accounts provide.