This post will be the last one in my mini-series “A Similarity Space for Shapes” about joint work with Margit Scheibel. So far, I have described the overall data set, the correlation between distances and dissimilarities, and the well-shapedness of conceptual regions. Today, I will finally take a look at interpretable directions in this similarity space.
Today, I would like to continue my little series about recent joint work with Margit Scheibel on a psychologically grounded similarity space for shapes. In my first post, I outlined the data set we worked with, and in my second post, we investigated how well the dissimilarity ratings are reflected by distances in the similarity spaces. Today, I’m going to use the categories from our data set to analyze whether conceptual regions are well-formed.
In one of my last blog posts, I have introduced a data set of shapes which I use to extract similarity spaces for the shape domain. As stated at the end of that blog post, I want to analyze these similarity spaces based on three predictions of the conceptual spaces framework: The representation of dissimilarities as distances, the presence of small, non-overlapping convex regions, and the presence of interpretable directions. Today, I will focus on the first of these predictions. More specifically, we will compute the correlation between distances in the MDS-spaces to the original dissimilarities and compare this to three baselines. This will help us to see how efficiently the similarity spaces represent shape similarity.
As already mentioned earlier, I want to validate my hybrid proposal for obtaining the dimensions of a conceptual space in a second study, which focuses on the domain of shapes. Today I will start reporting on joint work with Margit Scheibel on obtaining a similarity space for the shape domain based on psychological data. This is the first step of the proposed hybrid procedure and will be followed by training a neural network. But for now, let’s focus on obtaining the similarity spaces.
This blog post closes the “A Hybrid Way: Reloaded” mini-series. So far, I have analyzed the MDS solutions in part 1 and investigated first regression results in part 2 (with respect to the effects of feature space, correct vs. shuffled targets, and regularization). Today, I want to analyze what happens if we use different MDS algorithms for constructing the similarity spaces and to what extent our regression results depend on the number of dimensions in the similarity space.