November 21, 2012 10:00 - 10:30
We calculated the dimensionality of neural object representations using a large set of cells recorded in anterior inferotemporal cortex of macaque monkeys, stimulated with photographs of objects. Two unrelated dimensionality-estimation methods were used, the first being the Grassberger-Procaccia correlation dimension algorithm, and the second based on eigenvalues from a principal components analysis of the data. Both methods yielded similar dimensionality estimates of 40 and 52 respectively. However, the dimensionality estimate was found to be dependent on data set size, increasing as either the number of recorded neurons increased or the number of stimulus images increased. Using bootstrap resampling of the data we constructed a 2D function of dimensionality as a function of both number of neurons and images. Then we extrapolated the function to estimate the asymptotic dimensionality as data set size approached infinity. This yielded a final dimensionality estimate of approximately 200 for neural representations of objects in inferotemporal cortex. The dimensionality of the neural representation was much lower than the dimensionality of the physical stimuli, which was estimated as being approximately 800 using the same methods. Estimating the dimensionality of object representations provides a first step for a quantitative characterization of object space in general.
- BSI Private Event
- Sidney Lehky