Poster #5 - Learning order affects internal representation structure: A computational model of teaching in children's learning

• In Event: 2-055 - Poster Session 06
  In Poster Session: PS 06 Section - Attention, Learning, Memory

Integrative Statement

Children typically learn tasks in the order of increasing complexity. Structured teaching helps them to acquire new skills faster and to stay motivated. What is intuitively clear has been summarized in the concept of the Zone of Proximal Development (ZPD) (Vygotsky 1987), which suggests that for a medium level of task difficulty, the child's progress is maximal, whereas too simple or too difficult tasks impede learning progress. In this concept, the teacher plays an important role, providing the child with assistance and a structured environment. This scaffolding enables the child to solve tasks which are just beyond its current capacity (Wood et al. 1976).

In computational models, learning also is influenced by how training data are presented. Although there are significant differences to how children learn, analogies can be drawn, for instance, in terms of the order in which training data are presented. Computational models make it possible to investigate the internal mechanisms which cause behavioral phenomena. Here, we present a computational model to understand the influence of teaching order, with the aim to acquire new insights about the underlying causes of why children require structured teaching.

For this purpose, we trained recurrent neural networks (Murata et al. 2013) to draw four different shapes: ellipses (easy shapes) and "eight" shapes (more complex shapes), each at two different positions (see Figure 1). The networks are trained in epochs which correspond to learning progress.

We investigate the influence of teaching order by comparing different conditions of how training data are presented to the network: A basic learning condition where all shapes are presented to the network from the beginning, and several learning conditions where shapes are presented successively, adding a new shape every 1000 epochs. Either easier shapes or more difficult shapes were trained first.

All networks, regardless of teaching order, were able to generate the shapes with a small error after 6000 epochs (Figure 2a). However, there were significant differences in the quality of the internal representations that the networks developed during training (Figure 2b). This quality can be measured after training by testing how well the activation of the network's neurons reflect differences and similarities of the training data. In recurrent neural networks, the quality of the internal representation of a model is a good indicator for the network's capability to generalize to new tasks which were not explicitly trained (Tani et al. 2004). In accordance with the suggestion of the ZPD, we found that the best representation quality is achieved if the simpler ellipse shapes are trained first.

This result suggests that teaching order might be important in children's learning because the structured presentation of training data facilitates the development of a well structured internal representation, such as the neural representation in the child's brain. Furthermore, as internal representation quality is linked to the model's generalization capability, the scaffolding which a caregiver provides might not only have an influence on the visible progress of the child, but might also in general enhance the child's general problem solving capabilities.

Authors

• Anja Philippsen, National Institute of Information and Communications Technology
  Presenting Author

• Timothée Anne, École Normale Suprieure de Rennes
  Non-Presenting Author

• Yukie Nagai, National Institute of Information and Communications Technology
  Non-Presenting Author