Mapping Images to Psychological Similarity Spaces Using Neural Networks.

Mapping Images to Psychological Similarity Spaces Using Neural Networks.


The cognitive framework of conceptual spaces bridges the gap between symbolic and subsymbolic AI by proposing an intermediate conceptual layer where knowledge is represented geometrically. There are two main approaches for obtaining the dimensions of this conceptual similarity space: using similarity ratings from psychological experiments and using machine learning techniques. In this paper, we propose a combination of both approaches by using psychologically derived similarity ratings to constrain the machine learning process. This way, a mapping from stimuli to conceptual spaces can be learned that is both supported by psychological data and allows generalization to unseen stimuli. The results of a first feasibility study support our proposed approach.

In 6th International Workshop on Artificial Intelligence and Cognition (AIC).

We have argued that the combination of neural networks with psychological similarity judgments offers a promising way of extracting a conceptual space from data. This can help to make the framework of conceptual spaces more viable for artificial intelligence: Our proposed approach can potentially provide a principled way of mapping sensory input to conceptual spaces, while still maintaining some psychological validity. The results of our feasibility study are encouraging and show that our approach is also feasible in practice.

In future work, we will implement and evaluate our proposal in more depth by exploring the remaining proposed network structures. We will conduct a psychological study on a subset of ImageNet to investigate whether we can achieve a better mapping performance if we use images from the same domain. The data used in our current study has the shortcoming of not separating different domains like color, shape, and size (as proposed by Gardenfors), but of treatingthem as a single space. In order to evaluate whether this difference impacts the performance of our proposal, we will apply our approach to a single domain (namely, shapes). Finally, we will also explore additional ways of evaluating the mapping performance (e.g., by comparing to the original similarity ratings).