Spontaneous biases enhance generalisation in the neonate brain

Abstract

The ability to use sparse evidence to produce adaptive responses in new contexts and to new stimuli (inductive generalisation) is central to biological and artificial intelligence. Young and inexperienced animals require very little evidence to generalise, raising the question of whether the neonate brain is evolutionarily prepared (predisposed) for generalisation. To understand the principles of spontaneous generalisation, we exposed neonate chicks to an artificial social partner of a specific colour, and measured generalisation by comparing responses to novel and familiar stimuli along either the red-yellow or the blue-green colour continuum. Generalisation responses were inconsistent with an unbiased model, showing biases such as asymmetrical generalisation gradients, faster learning for particular colours (red and blue over yellow and green), preferences for unfamiliar stimuli and different time courses in learning. Moreover, the chicks generalisation behaviour was consistent with a Bayesian theoretical model that explicitly incorporates predispositions as initial preferences and treats the learning process as an update of spontaneous preferences. These results show that neonate animals are evolutionarily prepared for generalisation, via biases that do not depend on experience, reinforcement or supervision. Biases that facilitate generalisation are tuned to distinctive features that are unusual in the natural environment, such as the red and blue colours. Predispositions facilitate or hinder learning in the inexperienced brain, determining how experience is used to update the likelihood of predictive models. Neonate animals use spontaneous biases to solve the problem of induction.