Infants are curious learners who drive their own cognitive development by imposing structure on their learning environment as they explore. Understanding the mechanisms by which infants structure their own learning is therefore critical to our understanding of development. Here we propose an explicit mechanism for intrinsically motivated information selection that maximizes learning. We first present a neurocomputational model of infant visual category learning, capturing existing empirical data on the role of environmental complexity on learning. Next we “set the model free”, allowing it to select its own stimuli based on a formalization of curiosity and three alternative selection mechanisms. We demonstrate that maximal learning emerges when the model is able to maximize stimulus novelty relative to its internal states, depending on the interaction across learning between the structure of the environment and the plasticity in the learner itself. We discuss the implications of this new curiosity mechanism for both existing computational models of reinforcement learning and for our understanding of this fundamental mechanism in early development.The overall goal of this award was to understand how babies learn when allowed to explore their environment based on their own curiosity, outside the constrained experimental setting typical of most research in early cognitive development. We were also interested in how this curiosity-based exploration might be influenced by language. This goal was approached in two ways: first using computational modelling to examine the potential learning mechanisms involved in curiosity; and second, experimentally, to develop a picture of what babies and toddlers do when engaged in curiosity-driven learning. In our computational work we developed the first model of babies curiosity-driven learning inspired by the mechanisms known to exist in the human brain. This model predicted that when allowed to freely choose what to learn from and when, young children should learn best from an environment which is neither too simple nor too complex; that is, medium difficulty should best support learning, and importantly, children should be able to generate this level of difficulty themselves without adults structuring their learning environment on their behalf. Our empirical work aimed to test the predictions from the model. In Study 1 we showed 12- and 28-month-old toddlers 2D image and recorded where they looked and for how long. Both groups of children generated patterns of looking which were of intermediate complexity (Twomey, Malem, Ke & Westermann, in prep.). In Study 2, we allowed 12-, 18- and 24-month-old infants to play freely with custom-designed, 3D printed categories of novel objects. Again, children of all ages generated explored the objects in an order which led to medium complexity (Ke, Westermann & Twomey, in prep[a]). This study also generated a video dataset from the 12-month-old participants showing their field of view (Ke, Westermann & Twomey, in prep[b]). This dataset will allow us to conduct fine-grained analyses of their how young children visually explore the object they’re playing with, linking the findings from Studies 1 and 2. Overall the empirical data support the predictions of the model, providing the first evidence that not only do infants learn best from intermediate difficulty input, but critically also that they are capable of generating this level of difficulty independently. Put differently, rather than passive learners or random explorers, infants are active learners who are capable of independently tailoring their learning environment in a way that best supports their own development.

Neurocomputational modelling (autoencode neural network)