AI systems struggle with abstract visual reasoning - looking at patterns and inferring rules, like solving Raven's Progressive Matrices. Humans do this effortlessly. A study in IEEE Transactions on Pattern Analysis and Machine Intelligence proposes that a concept from neuroscience - prediction error - could help close this gap.
What Are Prediction Errors?
In the brain, prediction errors signal the difference between what you expected and what actually happened. Dopamine neurons famously encode these errors, firing more when outcomes are better than expected and less when they're worse.
The researchers propose that prediction errors can serve as a unified mechanism for abstract reasoning in AI systems.
The Model
Their approach frames reasoning as prediction-and-matching. Given a sequence of patterns following some abstract rule, the model predicts what should come next, then compares that prediction to candidate answers. The discrepancy - the prediction error - determines the best answer.
This works for both supervised learning (with labeled training data) and self-supervised learning (learning from the structure of the data itself).
State-of-the-Art Performance
The prediction error models achieved state-of-the-art performance on a broad range of abstract reasoning datasets and task conditions. They outperformed existing approaches that use different computational strategies.
Emergent Biological Properties
Perhaps most intriguingly, the hierarchical prediction errors in the supervised model automatically decreased during training. This emergence mirrors what happens in biological learning - dopamine prediction error signals decrease as animals learn to predict outcomes.
The parallel suggests the model isn't just achieving good performance through arbitrary computation, but may be capturing something fundamental about how intelligent systems learn to reason.
Reference: Yang L, et al. (2025). Neural Prediction Errors as A Unified Cue for Abstract Visual Reasoning. IEEE Trans Pattern Anal Mach Intell. doi: 10.1109/TPAMI.2025.3623461 | PMID: 41115076
Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.