Latent structure in random sequences drives neural learning toward a rational bias

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Psychology

Background:

• Humans exhibit a bias against repeating elements and overuse alternations when generating random sequences, a phenomenon linked to the gambler's fallacy.
• Existing models struggle to fully explain the neural underpinnings of this deviation from true randomness in human sequence production.

Purpose of the Study:

• To investigate the neural basis of the gambler's fallacy bias in human random sequence generation.
• To develop and validate a biologically motivated neural model capable of replicating human biases in randomness perception.

Main Methods:

• A biologically motivated neural network model was developed to learn from prediction errors when exposed to random sequences.
• The model's internal representations and learned parameters were analyzed to understand the emergence of alternating patterns.
• The model's output bias-gain parameter was directly compared with existing Bayesian models fitted to human data.

Main Results:

• The neural model naturally developed a bias towards alternation after exposure to random sequences, mirroring human behavior.
• This emergent bias was attributed to the model's sensitivity to subtle statistical structures within the random sequences.
• The model's bias-gain parameter accurately fitted human data, validating its explanatory power for the gambler's fallacy.

Conclusions:

• The human tendency to avoid repetition and favor alternation in random sequences is not necessarily irrational but an adaptive response of an effective learning mechanism.
• This bias reflects sensitivity to underlying statistical regularities in random data, as demonstrated by the neural model's behavior.
• The findings offer a computational explanation for a common cognitive bias, linking learning mechanisms to the perception of randomness.