| |
Abstract:
Abstract: Purpose: To investigate biologically plausible
models of visual object representation. Methods: 1) Subjects
assigned similarity ratings from 1-9 to pairs of objects including
line drawings of faces and fish. Each set of objects was
parameterized along four dimensions that were unknown to the
subjects. We used multidimensional scaling (MDS) to model subjects'
internal representations. 2) Subjects learned (2-AFC) to categorize
the objects from part 1 into two classes linearly separable in the
objects' parameter space. Subsequently, subjects categorized
additional test objects, and these results were used to fit a
weighted prototype similarity model (WPSM) and a generalized
context model (GCM) based either on the objects' parameter space or
the MDS space. 3) We are also repeating these experiments in a 1.5T
magnet using fMRI-BOLD. Results and Conclusions: For the face
stimuli, the MDS space rarely (3/20 experiments) improved the
models' performance, implying that subjects' internal
representations are naturally similar to the original space. In
contrast, with the more abstract fish stimuli, the MDS space
improved performance in 3 of 6 experiments. For all object types,
the GCM reliably (24/26 experiments) outperformed the WPSM,
suggesting category-level abstraction is not necessary for accurate
classification. However, the GCM predicts that individual exemplars
are remembered, which appears to be at odds with neural models of
memory. We are currently investigating this by varying the number
of exemplars.
|
